Foundations and Trends
R

in Signal Processing
Vol. 7, Nos. 3–4 (2013) 197–387

c 2014 L. Deng and D. Yu
DOI: 10.1561/2000000039

Deep Learning: Methods and Applications

Li Deng Dong Yu
Microsoft Research Microsoft Research
One Microsoft Way One Microsoft Way
Redmond, WA 98052; USA Redmond, WA 98052; USA
[email protected] [email protected]
 Contents

1 Introduction 198
1.1 Definitions and background . . . . . . . . . . . . . . . . . 198
1.2 Organization of this monograph . . . . . . . . . . . . . . 202

2 Some Historical Context of Deep Learning 205

3 Three Classes of Deep Learning Networks 214

3.1 A three-way categorization . . . . . . . . . . . . . . . . . 214
3.2 Deep networks for unsupervised or generative learning . . . 216
3.3 Deep networks for supervised learning . . . . . . . . . . . 223
3.4 Hybrid deep networks . . . . . . . . . . . . . . . . . . . . 226

4 Deep Autoencoders — Unsupervised Learning 230

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 230
4.2 Use of deep autoencoders to extract speech features . . . 231
4.3 Stacked denoising autoencoders . . . . . . . . . . . . . . . 235
4.4 Transforming autoencoders . . . . . . . . . . . . . . . . . 239

5 Pre-Trained Deep Neural Networks — A Hybrid 241

5.1 Restricted Boltzmann machines . . . . . . . . . . . . . . . 241
5.2 Unsupervised layer-wise pre-training . . . . . . . . . . . . 245
5.3 Interfacing DNNs with HMMs . . . . . . . . . . . . . . . 248

ii
 iii

6 Deep Stacking Networks and Variants —

Supervised Learning 250
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.2 A basic architecture of the deep stacking network . . . . . 252
6.3 A method for learning the DSN weights . . . . . . . . . . 254
6.4 The tensor deep stacking network . . . . . . . . . . . . . . 255
6.5 The Kernelized deep stacking network . . . . . . . . . . . 257

7 Selected Applications in Speech and Audio Processing 262

7.1 Acoustic modeling for speech recognition . . . . . . . . . . 262
7.2 Speech synthesis . . . . . . . . . . . . . . . . . . . . . . . 286
7.3 Audio and music processing . . . . . . . . . . . . . . . . . 288

8 Selected Applications in Language

Modeling and Natural Language Processing 292
8.1 Language modeling . . . . . . . . . . . . . . . . . . . . . 293
8.2 Natural language processing . . . . . . . . . . . . . . . . . 299

9 Selected Applications in Information Retrieval 308

9.1 A brief introduction to information retrieval . . . . . . . . 308
9.2 SHDA for document indexing and retrieval . . . . . . . . . 310
9.3 DSSM for document retrieval . . . . . . . . . . . . . . . . 311
9.4 Use of deep stacking networks for information retrieval . . 317

10 Selected Applications in Object Recognition

and Computer Vision 320
10.1 Unsupervised or generative feature learning . . . . . . . . 321
10.2 Supervised feature learning and classification . . . . . . . . 324

11 Selected Applications in Multimodal

and Multi-task Learning 331
11.1 Multi-modalities: Text and image . . . . . . . . . . . . . . 332
11.2 Multi-modalities: Speech and image . . . . . . . . . . . . 336
11.3 Multi-task learning within the speech, NLP or image . . . 339
iv

12 Conclusion 343

References 349
Abstract

This monogrph provides an overview of general deep learning method-

ology and its applications to a variety of signal and information pro-
cessing tasks. The application areas are chosen with the following three
criteria in mind: (1) expertise or knowledge of the authors; (2) the
application areas that have already been transformed by the successful
use of deep learning technology, such as speech recognition and com-
puter vision; and (3) the application areas that have the potential to be
impacted significantly by deep learning and that have been experienc-
ing research growth, including natural language and text processing,
information retrieval, and multimodal information processing empow-
ered by multi-task deep learning.

L. Deng and D. Yu. Deep Learning: Methods and Applications. Foundations and
Trends

R
in Signal Processing, vol. 7, nos. 3–4, pp. 197–387, 2013.
DOI: 10.1561/2000000039.
 1
Introduction

1.1 Definitions and background

Since 2006, deep structured learning, or more commonly called deep

learning or hierarchical learning, has emerged as a new area of machine
learning research [20, 163]. During the past several years, the techniques
developed from deep learning research have already been impacting
a wide range of signal and information processing work within the
traditional and the new, widened scopes including key aspects of
machine learning and artificial intelligence; see overview articles in
[7, 20, 24, 77, 94, 161, 412], and also the media coverage of this progress
in [6, 237]. A series of workshops, tutorials, and special issues or con-
ference special sessions in recent years have been devoted exclusively
to deep learning and its applications to various signal and information
processing areas. These include:

• 2008 NIPS Deep Learning Workshop;

• 2009 NIPS Workshop on Deep Learning for Speech Recognition
and Related Applications;
• 2009 ICML Workshop on Learning Feature Hierarchies;

198
1.1. Definitions and background 199

• 2011 ICML Workshop on Learning Architectures, Representa-

tions, and Optimization for Speech and Visual Information Pro-
cessing;
• 2012 ICASSP Tutorial on Deep Learning for Signal and Informa-
tion Processing;
• 2012 ICML Workshop on Representation Learning;
• 2012 Special Section on Deep Learning for Speech and Language
Processing in IEEE Transactions on Audio, Speech, and Lan-
guage Processing (T-ASLP, January);
• 2010, 2011, and 2012 NIPS Workshops on Deep Learning and
Unsupervised Feature Learning;
• 2013 NIPS Workshops on Deep Learning and on Output Repre-
sentation Learning;
• 2013 Special Issue on Learning Deep Architectures in IEEE
Transactions on Pattern Analysis and Machine Intelligence
(T-PAMI, September).
• 2013 International Conference on Learning Representations;
• 2013 ICML Workshop on Representation Learning Challenges;
• 2013 ICML Workshop on Deep Learning for Audio, Speech, and
Language Processing;
• 2013 ICASSP Special Session on New Types of Deep Neural Net-
work Learning for Speech Recognition and Related Applications.

The authors have been actively involved in deep learning research and
in organizing or providing several of the above events, tutorials, and
editorials. In particular, they gave tutorials and invited lectures on
this topic at various places. Part of this monograph is based on their
tutorials and lecture material.
Before embarking on describing details of deep learning, let’s pro-
vide necessary definitions. Deep learning has various closely related
definitions or high-level descriptions:

• Definition 1 : A class of machine learning techniques that

exploit many layers of non-linear information processing for
200 Introduction

supervised or unsupervised feature extraction and transforma-

tion, and for pattern analysis and classification.
• Definition 2 : “A sub-field within machine learning that is based
on algorithms for learning multiple levels of representation in
order to model complex relationships among data. Higher-level
features and concepts are thus defined in terms of lower-level
ones, and such a hierarchy of features is called a deep architec-
ture. Most of these models are based on unsupervised learning of
representations.” (Wikipedia on “Deep Learning” around March
2012.)
• Definition 3 : “A sub-field of machine learning that is based
on learning several levels of representations, corresponding to a
hierarchy of features or factors or concepts, where higher-level
concepts are defined from lower-level ones, and the same lower-
level concepts can help to define many higher-level concepts. Deep
learning is part of a broader family of machine learning methods
based on learning representations. An observation (e.g., an image)
can be represented in many ways (e.g., a vector of pixels), but
some representations make it easier to learn tasks of interest (e.g.,
is this the image of a human face?) from examples, and research
in this area attempts to define what makes better representations
and how to learn them.” (Wikipedia on “Deep Learning” around
February 2013.)
• Definition 4 : “Deep learning is a set of algorithms in machine
learning that attempt to learn in multiple levels, correspond-
ing to different levels of abstraction. It typically uses artificial
neural networks. The levels in these learned statistical models
correspond to distinct levels of concepts, where higher-level con-
cepts are defined from lower-level ones, and the same lower-
level concepts can help to define many higher-level concepts.”
See Wikipedia http://en.wikipedia.org/wiki/Deep_learning on
“Deep Learning” as of this most recent update in October 2013.
• Definition 5 : “Deep Learning is a new area of Machine Learning
research, which has been introduced with the objective of moving
Machine Learning closer to one of its original goals: Artificial
1.1. Definitions and background 201

Intelligence. Deep Learning is about learning multiple levels of

representation and abstraction that help to make sense of data
such as images, sound, and text.” See https://github.com/lisa-
lab/DeepLearningTutorials

Note that the deep learning that we discuss in this monograph is

about learning with deep architectures for signal and information pro-
cessing. It is not about deep understanding of the signal or infor-
mation, although in many cases they may be related. It should also
be distinguished from the overloaded term in educational psychology:
“Deep learning describes an approach to learning that is character-
ized by active engagement, intrinsic motivation, and a personal search
for meaning.” http://www.blackwellreference.com/public/tocnode?id=
g9781405161251_chunk_g97814051612516_ss1-1
Common among the various high-level descriptions of deep learning
above are two key aspects: (1) models consisting of multiple layers
or stages of nonlinear information processing; and (2) methods for
supervised or unsupervised learning of feature representation at
successively higher, more abstract layers. Deep learning is in the
intersections among the research areas of neural networks, artificial
intelligence, graphical modeling, optimization, pattern recognition,
and signal processing. Three important reasons for the popularity
of deep learning today are the drastically increased chip processing
abilities (e.g., general-purpose graphical processing units or GPGPUs),
the significantly increased size of data used for training, and the recent
advances in machine learning and signal/information processing
research. These advances have enabled the deep learning methods
to effectively exploit complex, compositional nonlinear functions, to
learn distributed and hierarchical feature representations, and to make
effective use of both labeled and unlabeled data.
Active researchers in this area include those at University of
Toronto, New York University, University of Montreal, Stanford
University, Microsoft Research (since 2009), Google (since about
2011), IBM Research (since about 2011), Baidu (since 2012), Facebook
(since 2013), UC-Berkeley, UC-Irvine, IDIAP, IDSIA, University
College London, University of Michigan, Massachusetts Institute of
202 Introduction

Technology, University of Washington, and numerous other places; see

http://deeplearning.net/deep-learning-research-groups-and-labs/ for
a more detailed list. These researchers have demonstrated empirical
successes of deep learning in diverse applications of computer vision,
phonetic recognition, voice search, conversational speech recognition,
speech and image feature coding, semantic utterance classifica-
tion, natural language understanding, hand-writing recognition, audio
processing, information retrieval, robotics, and even in the analysis of
molecules that may lead to discovery of new drugs as reported recently
by [237].
In addition to the reference list provided at the end of this mono-
graph, which may be outdated not long after the publication of this
monograph, there are a number of excellent and frequently updated
reading lists, tutorials, software, and video lectures online at:

• http://deeplearning.net/reading-list/
• http://ufldl.stanford.edu/wiki/index.php/
UFLDL_Recommended_Readings
• http://www.cs.toronto.edu/∼hinton/
• http://deeplearning.net/tutorial/
• http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial

1.2 Organization of this monograph

The rest of the monograph is organized as follows:

In Section 2, we provide a brief historical account of deep learning,
mainly from the perspective of how speech recognition technology has
been hugely impacted by deep learning, and how the revolution got
started and has gained and sustained immense momentum.
In Section 3, a three-way categorization scheme for a majority of
the work in deep learning is developed. They include unsupervised,
supervised, and hybrid deep learning networks, where in the latter cat-
egory unsupervised learning (or pre-training) is exploited to assist the
subsequent stage of supervised learning when the final tasks pertain to
classification. The supervised and hybrid deep networks often have the
1.2. Organization of this monograph 203

same type of architectures or the structures in the deep networks, but

the unsupervised deep networks tend to have different architectures
from the others.
Sections 4–6 are devoted, respectively, to three popular types of
deep architectures, one from each of the classes in the three-way cat-
egorization scheme reviewed in Section 3. In Section 4, we discuss
in detail deep autoencoders as a prominent example of the unsuper-
vised deep learning networks. No class labels are used in the learning,
although supervised learning methods such as back-propagation are
cleverly exploited when the input signal itself, instead of any label
information of interest to possible classification tasks, is treated as the
“supervision” signal.
In Section 5, as a major example in the hybrid deep network cate-
gory, we present in detail the deep neural networks with unsupervised
and largely generative pre-training to boost the effectiveness of super-
vised training. This benefit is found critical when the training data
are limited and no other appropriate regularization approaches (i.e.,
dropout) are exploited. The particular pre-training method based on
restricted Boltzmann machines and the related deep belief networks
described in this section has been historically significant as it ignited
the intense interest in the early applications of deep learning to speech
recognition and other information processing tasks. In addition to this
retrospective review, subsequent development and different paths from
the more recent perspective are discussed.
In Section 6, the basic deep stacking networks and their several
extensions are discussed in detail, which exemplify the discrimina-
tive, supervised deep learning networks in the three-way classification
scheme. This group of deep networks operate in many ways that are
distinct from the deep neural networks. Most notably, they use target
labels in constructing each of many layers or modules in the overall
deep networks. Assumptions made about part of the networks, such as
linear output units in each of the modules, simplify the learning algo-
rithms and enable a much wider variety of network architectures to
be constructed and learned than the networks discussed in Sections 4
and 5.
204 Introduction

In Sections 7–11, we select a set of typical and successful applica-

tions of deep learning in diverse areas of signal and information process-
ing. In Section 7, we review the applications of deep learning to speech
recognition, speech synthesis, and audio processing. Subsections sur-
rounding the main subject of speech recognition are created based on
several prominent themes on the topic in the literature.
In Section 8, we present recent results of applying deep learning to
language modeling and natural language processing, where we highlight
the key recent development in embedding symbolic entities such as
words into low-dimensional, continuous-valued vectors.
Section 9 is devoted to selected applications of deep learning to
information retrieval including web search.
In Section 10, we cover selected applications of deep learning to
image object recognition in computer vision. The section is divided to
two main classes of deep learning approaches: (1) unsupervised feature
learning, and (2) supervised learning for end-to-end and joint feature
learning and classification.
Selected applications to multi-modal processing and multi-task
learning are reviewed in Section 11, divided into three categories
according to the nature of the multi-modal data as inputs to the deep
learning systems. For single-modality data of speech, text, or image,
a number of recent multi-task learning studies based on deep learning
methods are reviewed in the literature.
Finally, conclusions are given in Section 12 to summarize the mono-
graph and to discuss future challenges and directions.
This short monograph contains the material expanded from two
tutorials that the authors gave, one at APSIPA in October 2011 and
the other at ICASSP in March 2012. Substantial updates have been
made based on the literature up to January 2014 (including the mate-
rials presented at NIPS-2013 and at IEEE-ASRU-2013 both held in
December of 2013), focusing on practical aspects in the fast develop-
ment of deep learning research and technology during the interim years.
 2
Some Historical Context of Deep Learning

Until recently, most machine learning and signal processing techniques

had exploited shallow-structured architectures. These architectures
typically contain at most one or two layers of nonlinear feature transfor-
mations. Examples of the shallow architectures are Gaussian mixture
models (GMMs), linear or nonlinear dynamical systems, conditional
random fields (CRFs), maximum entropy (MaxEnt) models, support
vector machines (SVMs), logistic regression, kernel regression, multi-
layer perceptrons (MLPs) with a single hidden layer including extreme
learning machines (ELMs). For instance, SVMs use a shallow linear
pattern separation model with one or zero feature transformation layer
when the kernel trick is used or otherwise. (Notable exceptions are the
recent kernel methods that have been inspired by and integrated with
deep learning; e.g. [9, 53, 102, 377]). Shallow architectures have been
shown effective in solving many simple or well-constrained problems,
but their limited modeling and representational power can cause dif-
ficulties when dealing with more complicated real-world applications
involving natural signals such as human speech, natural sound and
language, and natural image and visual scenes.

205
206 Some Historical Context of Deep Learning

Human information processing mechanisms (e.g., vision and audi-

tion), however, suggest the need of deep architectures for extracting
complex structure and building internal representation from rich sen-
sory inputs. For example, human speech production and perception
systems are both equipped with clearly layered hierarchical structures
in transforming the information from the waveform level to the linguis-
tic level [11, 12, 74, 75]. In a similar vein, the human visual system is
also hierarchical in nature, mostly in the perception side but interest-
ingly also in the “generation” side [43, 126, 287]). It is natural to believe
that the state-of-the-art can be advanced in processing these types of
natural signals if efficient and effective deep learning algorithms can be
developed.
Historically, the concept of deep learning originated from artifi-
cial neural network research. (Hence, one may occasionally hear the
discussion of “new-generation neural networks.”) Feed-forward neural
networks or MLPs with many hidden layers, which are often referred
to as deep neural networks (DNNs), are good examples of the models
with a deep architecture. Back-propagation (BP), popularized in 1980s,
has been a well-known algorithm for learning the parameters of these
networks. Unfortunately BP alone did not work well in practice then
for learning networks with more than a small number of hidden layers
(see a review and analysis in [20, 129]. The pervasive presence of local
optima and other optimization challenges in the non-convex objective
function of the deep networks are the main source of difficulties in the
learning. BP is based on local gradient information, and starts usu-
ally at some random initial points. It often gets trapped in poor local
optima when the batch-mode or even stochastic gradient descent BP
algorithm is used. The severity increases significantly as the depth of
the networks increases. This difficulty is partially responsible for steer-
ing away most of the machine learning and signal processing research
from neural networks to shallow models that have convex loss func-
tions (e.g., SVMs, CRFs, and MaxEnt models), for which the global
optimum can be efficiently obtained at the cost of reduced modeling
power, although there had been continuing work on neural networks
with limited scale and impact (e.g., [42, 45, 87, 168, 212, 263, 304].
 207

The optimization difficulty associated with the deep models was

empirically alleviated when a reasonably efficient, unsupervised learn-
ing algorithm was introduced in the two seminar papers [163, 164].
In these papers, a class of deep generative models, called deep belief
network (DBN), was introduced. A DBN is composed of a stack of
restricted Boltzmann machines (RBMs). A core component of the
DBN is a greedy, layer-by-layer learning algorithm which optimizes
DBN weights at time complexity linear to the size and depth of the
networks. Separately and with some surprise, initializing the weights
of an MLP with a correspondingly configured DBN often produces
much better results than that with the random weights. As such,
MLPs with many hidden layers, or deep neural networks (DNN),
which are learned with unsupervised DBN pre-training followed by
back-propagation fine-tuning is sometimes also called DBNs in the
literature [67, 260, 258]. More recently, researchers have been more
careful in distinguishing DNNs from DBNs [68, 161], and when DBN
is used to initialize the training of a DNN, the resulting network is
sometimes called the DBN–DNN [161].
Independently of the RBM development, in 2006 two alternative,
non-probabilistic, non-generative, unsupervised deep models were pub-
lished. One is an autoencoder variant with greedy layer-wise training
much like the DBN training [28]. Another is an energy-based model
with unsupervised learning of sparse over-complete representations
[297]. They both can be effectively used to pre-train a deep neural
network, much like the DBN.
In addition to the supply of good initialization points, the DBN
comes with other attractive properties. First, the learning algorithm
makes effective use of unlabeled data. Second, it can be interpreted
as a probabilistic generative model. Third, the over-fitting problem,
which is often observed in the models with millions of parameters such
as DBNs, and the under-fitting problem, which occurs often in deep
networks, can be effectively alleviated by the generative pre-training
step. An insightful analysis on what kinds of speech information DBNs
can capture is provided in [259].
Using hidden layers with many neurons in a DNN significantly
improves the modeling power of the DNN and creates many closely
208 Some Historical Context of Deep Learning

optimal configurations. Even if parameter learning is trapped into a

local optimum, the resulting DNN can still perform quite well since
the chance of having a poor local optimum is lower than when a small
number of neurons are used in the network. Using deep and wide neu-
ral networks, however, would cast great demand to the computational
power during the training process and this is one of the reasons why it
is not until recent years that researchers have started exploring both
deep and wide neural networks in a serious manner.
Better learning algorithms and different nonlinearities also con-
tributed to the success of DNNs. Stochastic gradient descend (SGD)
algorithms are the most efficient algorithm when the training set is large
and redundant as is the case for most applications [39]. Recently, SGD is
shown to be effective for parallelizing over many machines with an asyn-
chronous mode [69] or over multiple GPUs through pipelined BP [49].
Further, SGD can often allow the training to jump out of local optima
due to the noisy gradients estimated from a single or a small batch of
samples. Other learning algorithms such as Hessian free [195, 238] or
Krylov subspace methods [378] have shown a similar ability.
For the highly non-convex optimization problem of DNN learn-
ing, it is obvious that better parameter initialization techniques will
lead to better models since optimization starts from these initial mod-
els. What was not obvious, however, is how to efficiently and effec-
tively initialize DNN parameters and how the use of large amounts of
training data can alleviate the learning problem until more recently
[28, 20, 100, 64, 68, 163, 164, 161, 323, 376, 414]. The DNN parameter
initialization technique that attracted the most attention is the unsu-
pervised pretraining technique proposed in [163, 164] discussed earlier.
The DBN pretraining procedure is not the only one that allows
effective initialization of DNNs. An alternative unsupervised approach
that performs equally well is to pretrain DNNs layer by layer by con-
sidering each pair of layers as a de-noising autoencoder regularized by
setting a random subset of the input nodes to zero [20, 376]. Another
alternative is to use contractive autoencoders for the same purpose by
favoring representations that are more robust to the input variations,
i.e., penalizing the gradient of the activities of the hidden units with
respect to the inputs [303]. Further, Ranzato et al. [294] developed the
 209

sparse encoding symmetric machine (SESM), which has a very similar

architecture to RBMs as building blocks of a DBN. The SESM may also
be used to effectively initialize the DNN training. In addition to unsu-
pervised pretraining using greedy layer-wise procedures [28, 164, 295],
the supervised pretraining, or sometimes called discriminative pretrain-
ing, has also been shown to be effective [28, 161, 324, 432] and in cases
where labeled training data are abundant performs better than the
unsupervised pretraining techniques. The idea of the discriminative
pretraining is to start from a one-hidden-layer MLP trained with the
BP algorithm. Every time when we want to add a new hidden layer we
replace the output layer with a randomly initialized new hidden and
output layer and train the whole new MLP (or DNN) using the BP
algorithm. Different from the unsupervised pretraining techniques, the
discriminative pretraining technique requires labels.
Researchers who apply deep learning to speech and vision analyzed
what DNNs capture in speech and images. For example, [259] applied
a dimensionality reduction method to visualize the relationship among
the feature vectors learned by the DNN. They found that the DNN’s
hidden activity vectors preserve the similarity structure of the feature
vectors at multiple scales, and that this is especially true for the fil-
terbank features. A more elaborated visualization method, based on
a top-down generative process in the reverse direction of the classi-
fication network, was recently developed by Zeiler and Fergus [436]
for examining what features the deep convolutional networks capture
from the image data. The power of the deep networks is shown to
be their ability to extract appropriate features and do discrimination
jointly [210].
As another way to concisely introduce the DNN, we can review the
history of artificial neural networks using a “hype cycle,” which is a
graphic representation of the maturity, adoption and social applica-
tion of specific technologies. The 2012 version of the hype cycles graph
compiled by Gartner is shown in Figure 2.1. It intends to show how
a technology or application will evolve over time (according to five
phases: technology trigger, peak of inflated expectations, trough of dis-
illusionment, slope of enlightenment, and plateau of production), and
to provide a source of insight to manage its deployment.
210 Some Historical Context of Deep Learning

Figure 2.1: Gartner hyper cycle graph representing five phases of a technology
(http://en.wikipedia.org/wiki/Hype_cycle).

Applying the Gartner hyper cycle to the artificial neural network

development, we created Figure 2.2 to align different generations of
the neural network with the various phases designated in the hype
cycle. The peak activities (“expectations” or “media hype” on the ver-
tical axis) occurred in late 1980s and early 1990s, corresponding to the
height of what is often referred to as the “second generation” of neu-
ral networks. The deep belief network (DBN) and a fast algorithm for
training it were invented in 2006 [163, 164]. When the DBN was used
to initialize the DNN, the learning became highly effective and this has
inspired the subsequent fast growing research (“enlightenment” phase
shown in Figure 2.2). Applications of the DBN and DNN to industry-
scale speech feature extraction and speech recognition started in 2009
when leading academic and industrial researchers with both deep learn-
ing and speech expertise collaborated; see reviews in [89, 161]. This
collaboration fast expanded the work of speech recognition using deep
learning methods to increasingly larger successes [94, 161, 323, 414],
 211

Figure 2.2: Applying Gartner hyper cycle graph to analyzing the history of artificial
neural network technology (We thank our colleague John Platt during 2012 for
bringing this type of “Hyper Cycle” graph to our attention for concisely analyzing
the neural network history).

many of which will be covered in the remainder of this monograph.

The height of the “plateau of productivity” phase, not yet reached in
our opinion, is expected to be higher than that in the stereotypical
curve (circled with a question mark in Figure 2.2), and is marked by
the dashed line that moves straight up.
We show in Figure 2.3 the history of speech recognition, which
has been compiled by NIST, organized by plotting the word error rate
(WER) as a function of time for a number of increasingly difficult
speech recognition tasks. Note all WER results were obtained using the
GMM–HMM technology. When one particularly difficult task (Switch-
board) is extracted from Figure 2.3, we see a flat curve over many
years using the GMM–HMM technology but after the DNN technology
is used the WER drops sharply (marked by the red star in Figure 2.4).
212 Some Historical Context of Deep Learning

Figure 2.3: The famous NIST plot showing the historical speech recognition error
rates achieved by the GMM-HMM approach for a number of increasingly difficult
speech recognition tasks. Data source: http://itl.nist.gov/iad/mig/publications/
ASRhistory/index.html

Figure 2.4: Extracting WERs of one task from Figure 2.3 and adding the signifi-
cantly lower WER (marked by the star) achieved by the DNN technology.
 213

In the next section, an overview is provided on the various architec-

tures of deep learning, followed by more detailed expositions of a few
widely studied architectures and methods and by selected applications
in signal and information processing including speech and audio, natu-
ral language, information retrieval, vision, and multi-modal processing.
 3
Three Classes of Deep Learning Networks

3.1 A three-way categorization

As described earlier, deep learning refers to a rather wide class of

machine learning techniques and architectures, with the hallmark
of using many layers of non-linear information processing that are
hierarchical in nature. Depending on how the architectures and tech-
niques are intended for use, e.g., synthesis/generation or recognition/
classification, one can broadly categorize most of the work in this area
into three major classes:

1. Deep networks for unsupervised or generative learn-

ing, which are intended to capture high-order correlation of the
observed or visible data for pattern analysis or synthesis purposes
when no information about target class labels is available. Unsu-
pervised feature or representation learning in the literature refers
to this category of the deep networks. When used in the genera-
tive mode, may also be intended to characterize joint statistical
distributions of the visible data and their associated classes when
available and being treated as part of the visible data. In the

214
3.1. A three-way categorization 215

latter case, the use of Bayes rule can turn this type of generative
networks into a discriminative one for learning.
2. Deep networks for supervised learning, which are intended
to directly provide discriminative power for pattern classifica-
tion purposes, often by characterizing the posterior distributions
of classes conditioned on the visible data. Target label data are
always available in direct or indirect forms for such supervised
learning. They are also called discriminative deep networks.
3. Hybrid deep networks, where the goal is discrimination which
is assisted, often in a significant way, with the outcomes of genera-
tive or unsupervised deep networks. This can be accomplished by
better optimization or/and regularization of the deep networks
in category (2). The goal can also be accomplished when discrim-
inative criteria for supervised learning are used to estimate the
parameters in any of the deep generative or unsupervised deep
networks in category (1) above.
Note the use of “hybrid” in (3) above is different from that used
sometimes in the literature, which refers to the hybrid systems for
speech recognition feeding the output probabilities of a neural network
into an HMM [17, 25, 42, 261].
By the commonly adopted machine learning tradition (e.g.,
Chapter 28 in [264], and Reference [95], it may be natural to just clas-
sify deep learning techniques into deep discriminative models (e.g., deep
neural networks or DNNs, recurrent neural networks or RNNs, convo-
lutional neural networks or CNNs, etc.) and generative/unsupervised
models (e.g., restricted Boltzmann machine or RBMs, deep belief
networks or DBNs, deep Boltzmann machines (DBMs), regularized
autoencoders, etc.). This two-way classification scheme, however,
misses a key insight gained in deep learning research about how gener-
ative or unsupervised-learning models can greatly improve the training
of DNNs and other deep discriminative or supervised-learning mod-
els via better regularization or optimization. Also, deep networks for
unsupervised learning may not necessarily need to be probabilistic or be
able to meaningfully sample from the model (e.g., traditional autoen-
coders, sparse coding networks, etc.). We note here that more recent
216 Three Classes of Deep Learning Networks

studies have generalized the traditional denoising autoencoders so that

they can be efficiently sampled from and thus have become genera-
tive models [5, 24, 30]. Nevertheless, the traditional two-way classifi-
cation indeed points to several key differences between deep networks
for unsupervised and supervised learning. Compared between the two,
deep supervised-learning models such as DNNs are usually more effi-
cient to train and test, more flexible to construct, and more suitable for
end-to-end learning of complex systems (e.g., no approximate inference
and learning such as loopy belief propagation). On the other hand, the
deep unsupervised-learning models, especially the probabilistic gener-
ative ones, are easier to interpret, easier to embed domain knowledge,
easier to compose, and easier to handle uncertainty, but they are typi-
cally intractable in inference and learning for complex systems. These
distinctions are retained also in the proposed three-way classification
which is hence adopted throughout this monograph.
Below we review representative work in each of the above three
categories, where several basic definitions are summarized in Table 3.1.
Applications of these deep architectures, with varied ways of learn-
ing including supervised, unsupervised, or hybrid, are deferred to Sec-
tions 7–11.

3.2 Deep networks for unsupervised or generative learning

Unsupervised learning refers to no use of task specific supervision infor-

mation (e.g., target class labels) in the learning process. Many deep net-
works in this category can be used to meaningfully generate samples by
sampling from the networks, with examples of RBMs, DBNs, DBMs,
and generalized denoising autoencoders [23], and are thus generative
models. Some networks in this category, however, cannot be easily sam-
pled, with examples of sparse coding networks and the original forms
of deep autoencoders, and are thus not generative in nature.
Among the various subclasses of generative or unsupervised deep
networks, the energy-based deep models are the most common [28, 20,
213, 268]. The original form of the deep autoencoder [28, 100, 164],
which we will give more detail about in Section 4, is a typical example
3.2. Deep networks for unsupervised or generative learning 217

Table 3.1: Basic deep learning terminologies.

Deep Learning: a class of machine learning techniques, where many

layers of information processing stages in hierarchical supervised
architectures are exploited for unsupervised feature learning and for
pattern analysis/classification. The essence of deep learning is to
compute hierarchical features or representations of the observational
data, where the higher-level features or factors are defined from
lower-level ones. The family of deep learning methods have been
growing increasingly richer, encompassing those of neural networks,
hierarchical probabilistic models, and a variety of unsupervised and
supervised feature learning algorithms.
Deep belief network (DBN): probabilistic generative models
composed of multiple layers of stochastic, hidden variables. The top
two layers have undirected, symmetric connections between them.
The lower layers receive top-down, directed connections from the
layer above.
Boltzmann machine (BM): a network of symmetrically connected,
neuron-like units that make stochastic decisions about whether to be
on or off.
Restricted Boltzmann machine (RBM): a special type of BM
consisting of a layer of visible units and a layer of hidden units with
no visible-visible or hidden-hidden connections.
Deep neural network (DNN): a multilayer perceptron with many
hidden layers, whose weights are fully connected and are often
(although not always) initialized using either an unsupervised or a
supervised pretraining technique. (In the literature prior to 2012, a
DBN was often used incorrectly to mean a DNN.)
Deep autoencoder: a “discriminative” DNN whose output targets
are the data input itself rather than class labels; hence an
unsupervised learning model. When trained with a denoising
criterion, a deep autoencoder is also a generative model and can be
sampled from.
(Continued)
218 Three Classes of Deep Learning Networks

Table 3.1: (Continued)

Distributed representation: an internal representation of the

observed data in such a way that they are modeled as being explained
by the interactions of many hidden factors. A particular factor
learned from configurations of other factors can often generalize well
to new configurations. Distributed representations naturally occur in
a “connectionist” neural network, where a concept is represented by a
pattern of activity across a number of units and where at the same
time a unit typically contributes to many concepts. One key
advantage of such many-to-many correspondence is that they provide
robustness in representing the internal structure of the data in terms
of graceful degradation and damage resistance. Another key
advantage is that they facilitate generalizations of concepts and
relations, thus enabling reasoning abilities.

of this unsupervised model category. Most other forms of deep autoen-

coders are also unsupervised in nature, but with quite different prop-
erties and implementations. Examples are transforming autoencoders
[160], predictive sparse coders and their stacked version, and de-noising
autoencoders and their stacked versions [376].
Specifically, in de-noising autoencoders, the input vectors are first
corrupted by, for example, randomly selecting a percentage of the
inputs and setting them to zeros or adding Gaussian noise to them.
Then the parameters are adjusted for the hidden encoding nodes to
reconstruct the original, uncorrupted input data using criteria such as
mean square reconstruction error and KL divergence between the orig-
inal inputs and the reconstructed inputs. The encoded representations
transformed from the uncorrupted data are used as the inputs to the
next level of the stacked de-noising autoencoder.
Another prominent type of deep unsupervised models with genera-
tive capability is the deep Boltzmann machine or DBM [131, 315, 316,
348]. A DBM contains many layers of hidden variables, and has no con-
nections between the variables within the same layer. This is a special
case of the general Boltzmann machine (BM), which is a network of
3.2. Deep networks for unsupervised or generative learning 219

symmetrically connected units that are on or off based on a stochastic

mechanism. While having a simple learning algorithm, the general BMs
are very complex to study and very slow to train. In a DBM, each layer
captures complicated, higher-order correlations between the activities
of hidden features in the layer below. DBMs have the potential of learn-
ing internal representations that become increasingly complex, highly
desirable for solving object and speech recognition problems. Further,
the high-level representations can be built from a large supply of unla-
beled sensory inputs and very limited labeled data can then be used to
only slightly fine-tune the model for a specific task at hand.
When the number of hidden layers of DBM is reduced to one, we
have restricted Boltzmann machine (RBM). Like DBM, there are no
hidden-to-hidden and no visible-to-visible connections in the RBM. The
main virtue of RBM is that via composing many RBMs, many hidden
layers can be learned efficiently using the feature activations of one
RBM as the training data for the next. Such composition leads to deep
belief network (DBN), which we will describe in more detail, together
with RBMs, in Section 5.
The standard DBN has been extended to the factored higher-order
Boltzmann machine in its bottom layer, with strong results obtained
for phone recognition [64] and for computer vision [296]. This model,
called the mean-covariance RBM or mcRBM, recognizes the limitation
of the standard RBM in its ability to represent the covariance structure
of the data. However, it is difficult to train mcRBMs and to use them
at the higher levels of the deep architecture. Further, the strong results
published are not easy to reproduce. In the architecture described by
Dahl et al. [64], the mcRBM parameters in the full DBN are not fine-
tuned using the discriminative information, which is used for fine tuning
the higher layers of RBMs, due to the high computational cost. Subse-
quent work showed that when speaker adapted features are used, which
remove more variability in the features, mcRBM was not helpful [259].
Another representative deep generative network that can be used
for unsupervised (as well as supervised) learning is the sum–product
network or SPN [125, 289]. An SPN is a directed acyclic graph with
the observed variables as leaves, and with sum and product operations
as internal nodes in the deep network. The “sum” nodes give mixture
220 Three Classes of Deep Learning Networks

models, and the “product” nodes build up the feature hierarchy. Prop-
erties of “completeness” and “consistency” constrain the SPN in a desir-
able way. The learning of SPNs is carried out using the EM algorithm
together with back-propagation. The learning procedure starts with a
dense SPN. It then finds an SPN structure by learning its weights,
where zero weights indicate removed connections. The main difficulty
in learning SPNs is that the learning signal (i.e., the gradient) quickly
dilutes when it propagates to deep layers. Empirical solutions have been
found to mitigate this difficulty as reported in [289]. It was pointed
out in that early paper that despite the many desirable generative
properties in the SPN, it is difficult to fine tune the parameters using
the discriminative information, limiting its effectiveness in classifica-
tion tasks. However, this difficulty has been overcome in the subse-
quent work reported in [125], where an efficient BP-style discriminative
training algorithm for SPN was presented. Importantly, the standard
gradient descent, based on the derivative of the conditional likelihood,
suffers from the same gradient diffusion problem well known in the
regular DNNs. The trick to alleviate this problem in learning SPNs
is to replace the marginal inference with the most probable state of
the hidden variables and to propagate gradients through this “hard”
alignment only. Excellent results on small-scale image recognition tasks
were reported by Gens and Domingo [125].
Recurrent neural networks (RNNs) can be considered as another
class of deep networks for unsupervised (as well as supervised) learning,
where the depth can be as large as the length of the input data sequence.
In the unsupervised learning mode, the RNN is used to predict the data
sequence in the future using the previous data samples, and no addi-
tional class information is used for learning. The RNN is very powerful
for modeling sequence data (e.g., speech or text), but until recently
they had not been widely used partly because they are difficult to train
to capture long-term dependencies, giving rise to gradient vanishing or
gradient explosion problems which were known in early 1990s [29, 167].
These problems can now be dealt with more easily [24, 48, 85, 280].
Recent advances in Hessian-free optimization [238] have also partially
overcome this difficulty using approximated second-order information
or stochastic curvature estimates. In the more recent work [239], RNNs
3.2. Deep networks for unsupervised or generative learning 221

that are trained with Hessian-free optimization are used as a genera-

tive deep network in the character-level language modeling tasks, where
gated connections are introduced to allow the current input characters
to predict the transition from one latent state vector to the next. Such
generative RNN models are demonstrated to be well capable of gener-
ating sequential text characters. More recently, Bengio et al. [22] and
Sutskever [356] have explored variations of stochastic gradient descent
optimization algorithms in training generative RNNs and shown that
these algorithms can outperform Hessian-free optimization methods.
Mikolov et al. [248] have reported excellent results on using RNNs for
language modeling. Most recently, Mesnil et al. [242] and Yao et al.
[403] reported the success of RNNs in spoken language understanding.
We will review this set of work in Section 8.
There has been a long history in speech recognition research
where human speech production mechanisms are exploited to con-
struct dynamic and deep structure in probabilistic generative models;
for a comprehensive review, see the monograph by Deng [76]. Specif-
ically, the early work described in [71, 72, 83, 84, 99, 274] generalized
and extended the conventional shallow and conditionally independent
HMM structure by imposing dynamic constraints, in the form of poly-
nomial trajectory, on the HMM parameters. A variant of this approach
has been more recently developed using different learning techniques
for time-varying HMM parameters and with the applications extended
to speech recognition robustness [431, 416]. Similar trajectory HMMs
also form the basis for parametric speech synthesis [228, 326, 439, 438].
Subsequent work added a new hidden layer into the dynamic model to
explicitly account for the target-directed, articulatory-like properties in
human speech generation [45, 73, 74, 83, 96, 75, 90, 231, 232, 233, 251,
282]. More efficient implementation of this deep architecture with hid-
den dynamics is achieved with non-recursive or finite impulse response
(FIR) filters in more recent studies [76, 107, 105]. The above deep-
structured generative models of speech can be shown as special cases
of the more general dynamic network model and even more general
dynamic graphical models [35, 34]. The graphical models can comprise
many hidden layers to characterize the complex relationship between
the variables in speech generation. Armed with powerful graphical
222 Three Classes of Deep Learning Networks

modeling tool, the deep architecture of speech has more recently been
successfully applied to solve the very difficult problem of single-channel,
multi-talker speech recognition, where the mixed speech is the visible
variable while the un-mixed speech becomes represented in a new hid-
den layer in the deep generative architecture [301, 391]. Deep generative
graphical models are indeed a powerful tool in many applications due
to their capability of embedding domain knowledge. However, they are
often used with inappropriate approximations in inference, learning,
prediction, and topology design, all arising from inherent intractability
in these tasks for most real-world applications. This problem has been
addressed in the recent work of Stoyanov et al. [352], which provides
an interesting direction for making deep generative graphical models
potentially more useful in practice in the future. An even more drastic
way to deal with this intractability was proposed recently by Bengio
et al. [30], where the need to marginalize latent variables is avoided
altogether.
The standard statistical methods used for large-scale speech recog-
nition and understanding combine (shallow) hidden Markov models
for speech acoustics with higher layers of structure representing dif-
ferent levels of natural language hierarchy. This combined hierarchical
model can be suitably regarded as a deep generative architecture, whose
motivation and some technical detail may be found in Section 7 of the
recent monograph [200] on “Hierarchical HMM” or HHMM. Related
models with greater technical depth and mathematical treatment can
be found in [116] for HHMM and [271] for Layered HMM. These early
deep models were formulated as directed graphical models, missing the
key aspect of “distributed representation” embodied in the more recent
deep generative networks of the DBN and DBM discussed earlier in this
chapter. Filling in this missing aspect would help improve these gener-
ative models.
Finally, dynamic or temporally recursive generative models based
on neural network architectures can be found in [361] for human motion
modeling, and in [344, 339] for natural language and natural scene pars-
ing. The latter model is particularly interesting because the learning
algorithms are capable of automatically determining the optimal model
structure. This contrasts with other deep architectures such as DBN
3.3. Deep networks for supervised learning 223

where only the parameters are learned while the architectures need to
be pre-defined. Specifically, as reported in [344], the recursive struc-
ture commonly found in natural scene images and in natural language
sentences can be discovered using a max-margin structure prediction
architecture. It is shown that the units contained in the images or sen-
tences are identified, and the way in which these units interact with
each other to form the whole is also identified.

3.3 Deep networks for supervised learning

Many of the discriminative techniques for supervised learning in signal

and information processing are shallow architectures such as HMMs
[52, 127, 147, 186, 188, 290, 394, 418] and conditional random fields
(CRFs) [151, 155, 281, 400, 429, 446]. A CRF is intrinsically a shal-
low discriminative architecture, characterized by the linear relationship
between the input features and the transition features. The shallow
nature of the CRF is made most clear by the equivalence established
between the CRF and the discriminatively trained Gaussian models
and HMMs [148]. More recently, deep-structured CRFs have been devel-
oped by stacking the output in each lower layer of the CRF, together
with the original input data, onto its higher layer [428]. Various ver-
sions of deep-structured CRFs are successfully applied to phone recog-
nition [410], spoken language identification [428], and natural language
processing [428]. However, at least for the phone recognition task, the
performance of deep-structured CRFs, which are purely discrimina-
tive (non-generative), has not been able to match that of the hybrid
approach involving DBN, which we will take on shortly.
Morgan [261] gives an excellent review on other major existing
discriminative models in speech recognition based mainly on the tra-
ditional neural network or MLP architecture using back-propagation
learning with random initialization. It argues for the importance of
both the increased width of each layer of the neural networks and the
increased depth. In particular, a class of deep neural network models
forms the basis of the popular “tandem” approach [262], where the out-
put of the discriminatively learned neural network is treated as part
224 Three Classes of Deep Learning Networks

of the observation variable in HMMs. For some representative recent

work in this area, see [193, 283].
In more recent work of [106, 110, 218, 366, 377], a new deep learning
architecture, sometimes called deep stacking network (DSN), together
with its tensor variant [180, 181] and its kernel version [102], are
developed that all focus on discrimination with scalable, parallelizable,
block-wise learning relying on little or no generative component. We
will describe this type of discriminative deep architecture in detail in
Section 6.
As discussed in the preceding section, recurrent neural networks
(RNNs) have been used as a generative model; see also the neural pre-
dictive model [87] with a similar “generative” mechanism. RNNs can
also be used as a discriminative model where the output is a label
sequence associated with the input data sequence. Note that such dis-
criminative RNNs or sequence models were applied to speech a long
time ago with limited success. In [17], an HMM was trained jointly with
the neural networks, with a discriminative probabilistic training crite-
rion. In [304], a separate HMM was used to segment the sequence during
training, and the HMM was also used to transform the RNN classifi-
cation results into label sequences. However, the use of the HMM for
these purposes does not take advantage of the full potential of RNNs.
A set of new models and methods were proposed more recently
in [133, 134, 135, 136] that enable the RNNs themselves to perform
sequence classification while embedding the long-short-term memory
into the model, removing the need for pre-segmenting the training data
and for post-processing the outputs. Underlying this method is the idea
of interpreting RNN outputs as the conditional distributions over all
possible label sequences given the input sequences. Then, a differen-
tiable objective function can be derived to optimize these conditional
distributions over the correct label sequences, where the segmentation
of the data is performed automatically by the algorithm. The effective-
ness of this method has been demonstrated in handwriting recognition
tasks and in a small speech task [135, 136] to be discussed in more
detail in Section 7 of this monograph.
Another type of discriminative deep architecture is the convo-
lutional neural network (CNN), in which each module consists of
3.3. Deep networks for supervised learning 225

a convolutional layer and a pooling layer. These modules are often

stacked up with one on top of another, or with a DNN on top of it, to
form a deep model [212]. The convolutional layer shares many weights,
and the pooling layer subsamples the output of the convolutional layer
and reduces the data rate from the layer below. The weight sharing
in the convolutional layer, together with appropriately chosen pool-
ing schemes, endows the CNN with some “invariance” properties (e.g.,
translation invariance). It has been argued that such limited “invari-
ance” or equi-variance is not adequate for complex pattern recognition
tasks and more principled ways of handling a wider range of invariance
may be needed [160]. Nevertheless, CNNs have been found highly effec-
tive and been commonly used in computer vision and image recognition
[54, 55, 56, 57, 69, 198, 209, 212, 434]. More recently, with appropri-
ate changes from the CNN designed for image analysis to that taking
into account speech-specific properties, the CNN is also found effec-
tive for speech recognition [1, 2, 3, 81, 94, 312]. We will discuss such
applications in more detail in Section 7 of this monograph.
It is useful to point out that the time-delay neural network (TDNN)
[202, 382] developed for early speech recognition is a special case and
predecessor of the CNN when weight sharing is limited to one of the
two dimensions, i.e., time dimension, and there is no pooling layer. It
was not until recently that researchers have discovered that the time-
dimension invariance is less important than the frequency-dimension
invariance for speech recognition [1, 3, 81]. A careful analysis on the
underlying reasons is described in [81], together with a new strategy for
designing the CNN’s pooling layer demonstrated to be more effective
than all previous CNNs in phone recognition.
It is also useful to point out that the model of hierarchical tempo-
ral memory (HTM) [126, 143, 142] is another variant and extension of
the CNN. The extension includes the following aspects: (1) Time or
temporal dimension is introduced to serve as the “supervision” infor-
mation for discrimination (even for static images); (2) Both bottom-up
and top-down information flows are used, instead of just bottom-up in
the CNN; and (3) A Bayesian probabilistic formalism is used for fusing
information and for decision making.
226 Three Classes of Deep Learning Networks

Finally, the learning architecture developed for bottom-up,

detection-based speech recognition proposed in [214] and developed
further since 2004, notably in [330, 332, 427] using the DBN–DNN
technique, can also be categorized in the discriminative or supervised-
learning deep architecture category. There is no intent and mecha-
nism in this architecture to characterize the joint probability of data
and recognition targets of speech attributes and of the higher-level
phone and words. The most current implementation of this approach
is based on the DNN, or neural networks with many layers using back-
propagation learning. One intermediate neural network layer in the
implementation of this detection-based framework explicitly represents
the speech attributes, which are simplified entities from the “atomic”
units of speech developed in the early work of [101, 355]. The simpli-
fication lies in the removal of the temporally overlapping properties
of the speech attributes or articulatory-like features. Embedding such
more realistic properties in the future work is expected to improve the
accuracy of speech recognition further.

3.4 Hybrid deep networks

The term “hybrid” for this third category refers to the deep architecture
that either comprises or makes use of both generative and discrimina-
tive model components. In the existing hybrid architectures published
in the literature, the generative component is mostly exploited to help
with discrimination, which is the final goal of the hybrid architecture.
How and why generative modeling can help with discrimination can be
examined from two viewpoints [114]:

• The optimization viewpoint where generative models trained in

an unsupervised fashion can provide excellent initialization points
in highly nonlinear parameter estimation problems (The com-
monly used term of “pre-training” in deep learning has been intro-
duced for this reason); and/or
• The regularization perspective where the unsupervised-learning
models can effectively provide a prior on the set of functions
representable by the model.
3.4. Hybrid deep networks 227

The study reported in [114] provided an insightful analysis and exper-

imental evidence supporting both of the viewpoints above.
The DBN, a generative, deep network for unsupervised learning dis-
cussed in Section 3.2, can be converted to and used as the initial model
of a DNN for supervised learning with the same network structure,
which is further discriminatively trained or fine-tuned using the target
labels provided. When the DBN is used in this way we consider this
DBN–DNN model as a hybrid deep model, where the model trained
using unsupervised data helps to make the discriminative model effec-
tive for supervised learning. We will review details of the discriminative
DNN for supervised learning in the context of RBM/DBN generative,
unsupervised pre-training in Section 5.
Another example of the hybrid deep network is developed in [260],
where the DNN weights are also initialized from a generative DBN
but are further fine-tuned with a sequence-level discriminative crite-
rion, which is the conditional probability of the label sequence given
the input feature sequence, instead of the frame-level criterion of cross-
entropy commonly used. This can be viewed as a combination of the
static DNN with the shallow discriminative architecture of CRF. It can
be shown that such a DNN–CRF is equivalent to a hybrid deep architec-
ture of DNN and HMM whose parameters are learned jointly using the
full-sequence maximum mutual information (MMI) criterion between
the entire label sequence and the input feature sequence. A closely
related full-sequence training method designed and implemented for
much larger tasks is carried out more recently with success for a shallow
neural network [194] and for a deep one [195, 353, 374]. We note that
the origin of the idea for joint training of the sequence model (e.g., the
HMM) and of the neural network came from the early work of [17, 25],
where shallow neural networks were trained with small amounts of
training data and with no generative pre-training.
Here, it is useful to point out a connection between the above
pretraining/fine-tuning strategy associated with hybrid deep networks
and the highly popular minimum phone error (MPE) training technique
for the HMM (see [147, 290] for an overview). To make MPE training
effective, the parameters need to be initialized using an algorithm (e.g.,
Baum-Welch algorithm) that optimizes a generative criterion (e.g.,
228 Three Classes of Deep Learning Networks

maximum likelihood). This type of methods, which uses maximum-

likelihood trained parameters to assist in the discriminative HMM
training can be viewed as a “hybrid” approach to train the shallow
HMM model.
Along the line of using discriminative criteria to train parameters in
generative models as in the above HMM training example, we here dis-
cuss the same method applied to learning other hybrid deep networks.
In [203], the generative model of RBM is learned using the discrimina-
tive criterion of posterior class-label probabilities. Here the label vector
is concatenated with the input data vector to form the combined vis-
ible layer in the RBM. In this way, RBM can serve as a stand-alone
solution to classification problems and the authors derived a discrim-
inative learning algorithm for RBM as a shallow generative model. In
the more recent work by Ranzato et al. [298], the deep generative model
of DBN with gated Markov random field (MRF) at the lowest level is
learned for feature extraction and then for recognition of difficult image
classes including occlusions. The generative ability of the DBN facil-
itates the discovery of what information is captured and what is lost
at each level of representation in the deep model, as demonstrated in
[298]. A related study on using the discriminative criterion of empirical
risk to train deep graphical models can be found in [352].
A further example of hybrid deep networks is the use of generative
models of DBNs to pre-train deep convolutional neural networks (deep
CNNs) [215, 216, 217]. Like the fully connected DNN discussed ear-
lier, pre-training also helps to improve the performance of deep CNNs
over random initialization. Pre-training DNNs or CNNs using a set of
regularized deep autoencoders [24], including denoising autoencoders,
contractive autoencoders, and sparse autoencoders, is also a similar
example of the category of hybrid deep networks.
The final example given here for hybrid deep networks is based
on the idea and work of [144, 267], where one task of discrimination
(e.g., speech recognition) produces the output (text) that serves
as the input to the second task of discrimination (e.g., machine
translation). The overall system, giving the functionality of speech
translation — translating speech in one language into text in another
language — is a two-stage deep architecture consisting of both
3.4. Hybrid deep networks 229

generative and discriminative elements. Both models of speech

recognition (e.g., HMM) and of machine translation (e.g., phrasal
mapping and non-monotonic alignment) are generative in nature, but
their parameters are all learned for discrimination of the ultimate
translated text given the speech data. The framework described in
[144] enables end-to-end performance optimization in the overall deep
architecture using the unified learning framework initially published
in [147]. This hybrid deep learning approach can be applied to not
only speech translation but also all speech-centric and possibly other
information processing tasks such as speech information retrieval,
speech understanding, cross-lingual speech/text understanding and
retrieval, etc. (e.g., [88, 94, 145, 146, 366, 398]).
In the next three chapters, we will elaborate on three prominent
types of models for deep learning, one from each of the three classes
reviewed in this chapter. These are chosen to serve the tutorial purpose,
given their simplicity of the architectural and mathematical descrip-
tions. The three architectures described in the following three chapters
may not be interpreted as the most representative and influential work
in each of the three classes.
 4
Deep Autoencoders — Unsupervised Learning

This section and the next two will each select one prominent example
deep network for each of the three categories outlined in Section 3.
Here we begin with the category of the deep models designed mainly
for unsupervised learning.

4.1 Introduction

The deep autoencoder is a special type of the DNN (with no class

labels), whose output vectors have the same dimensionality as the input
vectors. It is often used for learning a representation or effective encod-
ing of the original data, in the form of input vectors, at hidden layers.
Note that the autoencoder is a nonlinear feature extraction method
without using class labels. As such, the features extracted aim at con-
serving and better representing information instead of performing clas-
sification tasks, although sometimes these two goals are correlated.
An autoencoder typically has an input layer which represents the
original data or input feature vectors (e.g., pixels in image or spec-
tra in speech), one or more hidden layers that represent the trans-
formed feature, and an output layer which matches the input layer for

230
4.2. Use of deep autoencoders to extract speech features 231

reconstruction. When the number of hidden layers is greater than one,

the autoencoder is considered to be deep. The dimension of the hidden
layers can be either smaller (when the goal is feature compression) or
larger (when the goal is mapping the feature to a higher-dimensional
space) than the input dimension.
An autoencoder is often trained using one of the many back-
propagation variants, typically the stochastic gradient descent method.
Though often reasonably effective, there are fundamental problems
when using back-propagation to train networks with many hidden
layers. Once the errors get back-propagated to the first few layers,
they become minuscule, and training becomes quite ineffective. Though
more advanced back-propagation methods help with this problem to
some degree, it still results in slow learning and poor solutions, espe-
cially with limited amounts of training data. As mentioned in the pre-
vious chapters, the problem can be alleviated by pre-training each layer
as a simple autoencoder [28, 163]. This strategy has been applied to
construct a deep autoencoder to map images to short binary code for
fast, content-based image retrieval, to encode documents (called seman-
tic hashing), and to encode spectrogram-like speech features which we
review below.

4.2 Use of deep autoencoders to extract speech features

Here we review a set of work, some of which was published in [100],

in developing an autoencoder for extracting binary speech codes from
the raw speech spectrogram data in an unsupervised manner (i.e., no
speech class labels). The discrete representations in terms of a binary
code extracted by this model can be used in speech information retrieval
or as bottleneck features for speech recognition.
A deep generative model of patches of spectrograms that con-
tain 256 frequency bins and 1, 3, 9, or 13 frames is illustrated in
Figure 4.1. An undirected graphical model called a Gaussian-Bernoulli
RBM is built that has one visible layer of linear variables with
Gaussian noise and one hidden layer of 500 to 3000 binary latent
variables. After learning the Gaussian-Bernoulli RBM, the activation
232 Deep Autoencoders — Unsupervised Learning

Figure 4.1: The architecture of the deep autoencoder used in [100] for extracting
binary speech codes from high-resolution spectrograms. [after [100], @Elsevier].

probabilities of its hidden units are treated as the data for training
another Bernoulli-Bernoulli RBM. These two RBM’s can then be com-
posed to form a deep belief net (DBN) in which it is easy to infer the
states of the second layer of binary hidden units from the input in a
single forward pass. The DBN used in this work is illustrated on the left
side of Figure 4.1, where the two RBMs are shown in separate boxes.
(See more detailed discussions on the RBM and DBN in Section 5).
The deep autoencoder with three hidden layers is formed by
“unrolling” the DBN using its weight matrices. The lower layers of
this deep autoencoder use the matrices to encode the input and the
upper layers use the matrices in reverse order to decode the input.
This deep autoencoder is then fine-tuned using error back-propagation
to minimize the reconstruction error, as shown on the right side of Fig-
ure 4.1. After learning is complete, any variable-length spectrogram
4.2. Use of deep autoencoders to extract speech features 233

can be encoded and reconstructed as follows. First, N consecutive

overlapping frames of 256-point log power spectra are each normalized
to zero-mean and unit-variance across samples per feature to provide
the input to the deep autoencoder. The first hidden layer then uses the
logistic function to compute real-valued activations. These real values
are fed to the next, coding layer to compute “codes.” The real-valued
activations of hidden units in the coding layer are quantized to be
either zero or one with 0.5 as the threshold. These binary codes are
then used to reconstruct the original spectrogram, where individual
fixed-frame patches are reconstructed first using the two upper layers
of network weights. Finally, the standard overlap-and-add technique in
signal processing is used to reconstruct the full-length speech spectro-
gram from the outputs produced by applying the deep autoencoder to
every possible window of N consecutive frames. We show some illus-
trative encoding and reconstruction examples below.

Figure 4.2: Top to Bottom: The ordinal spectrogram; reconstructions using input
window sized of N = 1, 3, 9, and 13 while forcing the coding units to take values of
zero one (i.e., a binary code) . [after [100], @Elsevier].
234 Deep Autoencoders — Unsupervised Learning

At the top of Figure 4.2 is the original, un-coded speech, followed

by the speech utterances reconstructed from the binary codes (zero
or one) at the 312 unit bottleneck code layer with encoding window
lengths of N = 1, 3, 9, and 13, respectively. The lower reconstruction
errors for N = 9 and N = 13 are clearly seen.
Encoding error of the deep autoencoder is qualitatively examined
in comparison with the more traditional codes via vector quantization
(VQ). Figure 4.3 shows various aspects of the encoding errors. At the
top is the original speech utterance’s spectrogram. The next two spec-
trograms are the blurry reconstruction from the 312-bit VQ and the
much more faithful reconstruction from the 312-bit deep autoencoder.
Coding errors from both coders, plotted as a function of time, are

Figure 4.3: Top to bottom: The original spectrogram from the test set; reconstruc-
tion from the 312-bit VQ coder; reconstruction from the 312-bit autoencoder; coding
errors as a function of time for the VQ coder (blue) and autoencoder (red); spec-
trogram of the VQ coder residual; spectrogram of the deep autoencoder’s residual.
[after [100], @ Elsevier].
4.3. Stacked denoising autoencoders 235

Figure 4.4: The original speech spectrogram and the reconstructed counterpart.
A total of 312 binary codes are with one for each single frame.

shown below the spectrograms, demonstrating that the autoencoder

(red curve) is producing lower errors than the VQ coder (blue curve)
throughout the entire span of the utterance. The final two spectrograms
show detailed coding error distributions over both time and frequency
bins.
Figures 4.4 to 4.10 show additional examples (unpublished) for the
original un-coded speech spectrograms and their reconstructions using
the deep autoencoder. They give a diverse number of binary codes for
either a single or three consecutive frames in the spectrogram samples.

4.3 Stacked denoising autoencoders

In early years of autoencoder research, the encoding layer had smaller

dimensions than the input layer. However, in some applications, it is
desirable that the encoding layer is wider than the input layer, in which
case techniques are needed to prevent the neural network from learning
the trivial identity mapping function. One of the reasons for using a
236 Deep Autoencoders — Unsupervised Learning

Figure 4.5: Same as Figure 4.4 but with a different TIMIT speech utterance.

Figure 4.6: The original speech spectrogram and the reconstructed counterpart.
A total of 936 binary codes are used for three adjacent frames.
4.3. Stacked denoising autoencoders 237

Figure 4.7: Same as Figure 4.6 but with a different TIMIT speech utterance.

Figure 4.8: Same as Figure 4.6 but with yet another TIMIT speech utterance.

higher dimension in the hidden or encoding layers than the input layer
is that it allows the autoencoder to capture a rich input distribution.
The trivial mapping problem discussed above can be prevented by
methods such as using sparseness constraints, or using the “dropout”
trick by randomly forcing certain values to be zero and thus introducing
distortions at the input data [376, 375] or at the hidden layers [166]. For
238 Deep Autoencoders — Unsupervised Learning

Figure 4.9: The original speech spectrogram and the reconstructed counterpart.
A total of 2000 binary codes with one for each single frame.

Figure 4.10: Same as Figure 4.9 but with a different TIMIT speech utterance.

example, in the stacked denoising autoencoder detailed in [376], random

noises are added to the input data. This serves several purposes. First,
by forcing the output to match the original undistorted input data the
model can avoid learning the trivial identity solution. Second, since
the noises are added randomly, the model learned would be robust to
the same kind of distortions in the test data. Third, since each distorted
4.4. Transforming autoencoders 239

input sample is different, it greatly increases the training set size and
thus can alleviate the overfitting problem.
It is interesting to note that when the encoding and decoding
weights are forced to be the transpose of each other, such denoising
autoencoder with a single sigmoidal hidden layer is strictly equiva-
lent to a particular Gaussian RBM, but instead of training it by the
technique of contrastive divergence (CD) or persistent CD, it is trained
by a score matching principle, where the score is defined as the deriva-
tive of the log-density with respect to the input [375]. Furthermore,
Alain and Bengio [5] generalized this result to any parameterization
of the encoder and decoder with squared reconstruction error and
Gaussian corruption noise. They show that as the amount of noise
approaches zero, such models estimate the true score of the underly-
ing data generating distribution. Finally, Bengio et al. [30] show that
any denoising autoencoder is a consistent estimator of the underly-
ing data generating distribution within some family of distributions.
This is true for any parameterization of the autoencoder, for any
type of information-destroying corruption process with no constraint
on the noise level except being positive, and for any reconstruction
loss expressed as a conditional log-likelihood. The consistency of the
estimator is achieved by associating the denoising autoencoder with
a Markov chain whose stationary distribution is the distribution esti-
mated by the model, and this Markov chain can be used to sample
from the denoising autoencoder.

4.4 Transforming autoencoders

The deep autoencoder described above can extract faithful codes for
feature vectors due to many layers of nonlinear processing. However, the
code extracted in this way is transformation-variant. In other words,
the extracted code would change in ways chosen by the learner when the
input feature vector is transformed. Sometimes, it is desirable to have
the code change predictably to reflect the underlying transformation-
invariant property of the perceived content. This is the goal of the
transforming autoencoder proposed in [162] for image recognition.
240 Deep Autoencoders — Unsupervised Learning

The building block of the transforming autoencoder is a “capsule,”

which is an independent sub-network that extracts a single parameter-
ized feature representing a single entity, be it visual or audio. A trans-
forming autoencoder receives both an input vector and a target output
vector, which is transformed from the input vector through a simple
global transformation mechanism; e.g., translation of an image and
frequency shift of speech (the latter due to the vocal tract length
difference). An explicit representation of the global transformation is
assumed known. The coding layer of the transforming autoencoder con-
sists of the outputs of several capsules.
During the training phase, the different capsules learn to extract
different entities in order to minimize the error between the final output
and the target.
In addition to the deep autoencoder architectures described here,
there are many other types of generative architectures in the literature,
all characterized by the use of data alone (i.e., free of classification
labels) to automatically derive higher-level features.
 5
Pre-Trained Deep Neural Networks — A Hybrid

In this section, we present the most widely used hybrid deep archi-
tecture — the pre-trained deep neural network (DNN), and discuss
the related techniques and building blocks including the RBM and
DBN. We discuss the DNN example here in the category of hybrid
deep networks before the examples in the category of deep networks for
supervised learning (Section 6). This is partly due to the natural flow
from the unsupervised learning models to the DNN as a hybrid model.
The discriminative nature of artificial neural networks for supervised
learning has been widely known, and thus would not be required for
understanding the hybrid nature of the DNN that uses unsupervised
pre-training to facilitate the subsequent discriminative fine tuning.
Part of the review in this chapter is based on recent publications in
[68, 161, 412].

5.1 Restricted Boltzmann machines

An RBM is a special type of Markov random field that has one layer of
(typically Bernoulli) stochastic hidden units and one layer of (typically
Bernoulli or Gaussian) stochastic visible or observable units. RBMs can

241
242 Pre-Trained Deep Neural Networks — A Hybrid

be represented as bipartite graphs, where all visible units are connected

to all hidden units, and there are no visible–visible or hidden–hidden
connections.
In an RBM, the joint distribution p(v, h; θ) over the visible units v
and hidden units h, given the model parameters θ, is defined in terms
of an energy function E(v, h; θ) of
exp(−E(v, h; θ))
p(v, h; θ) = ,
Z



where Z = v h exp(−E(v, h; θ)) is a normalization factor or parti-
tion function, and the marginal probability that the model assigns to
a visible vector v is


h exp(−E(v, h; θ))
p(v; θ) =
Z
For a Bernoulli (visible)-Bernoulli (hidden) RBM, the energy function
is defined as
I 
 J I
 J

E(v, h; θ) = − wij vi hj − bi vi − aj hj .
i=1 j=1 i=1 j=1

where wij represents the symmetric interaction term between visible

unit vi and hidden unit hj , bi and aj the bias terms, and I and J are
the numbers of visible and hidden units. The conditional probabilities
can be efficiently calculated as
 I 

p(hj = 1|v; θ) = σ wij vi + aj ,
i=1
 
J

p(vi = 1|h; θ) = σ  wij hj +bi  ,
j=1

where σ(x) = 1/(1 + exp(−x)).

Similarly, for a Gaussian (visible)-Bernoulli (hidden) RBM, the
energy is
I 
 J I J
1 
E(v, h; θ) = − wij vi hj − (vi − bi ) −
2
aj hj ,
i=1 j=1
2 i=1 j=1
5.1. Restricted Boltzmann machines 243

The corresponding conditional probabilities become

 I 

p(hj = 1|v; θ) = σ wij vi +aj ,
i=1
 
J

p(vi |h; θ) = N  wij hj + bi , 1 ,
j=1

where vi takes real values and follows a Gaussian distribution with

mean Jj=1 wij hj + bi and variance one. Gaussian-Bernoulli RBMs can
be used to convert real-valued stochastic variables to binary stochastic
variables, which can then be further processed using the Bernoulli-
Bernoulli RBMs.
The above discussion used two of the most common conditional
distributions for the visible data in the RBM — Gaussian (for
continuous-valued data) and binomial (for binary data). More general
types of distributions in the RBM can also be used. See [386] for the
use of general exponential-family distributions for this purpose.
Taking the gradient of the log likelihood log p(v; θ) we can derive
the update rule for the RBM weights as:

∆wij = Edata (vi hj ) − Emodel (vi hj ),

where Edata (vi hj ) is the expectation observed in the training set (with
hj sampled given vi according to the model), and Emodel (vi hj ) is that
same expectation under the distribution defined by the model. Unfor-
tunately, Emodel (vi hj ) is intractable to compute. The contrastive diver-
gence (CD) approximation to the gradient was the first efficient method
proposed to approximate this expected value, where Emodel (vi hj ) is
replaced by running the Gibbs sampler initialized at the data for one
or more steps. The steps in approximating Emodel (vi hj ) is summarized
as follows:

• Initialize v0 at data
• Sample h0 ∼ p(h|v0 )
• Sample v1 ∼ p(v|h0 )
• Sample h1 ∼ p(h|v1 )
244 Pre-Trained Deep Neural Networks — A Hybrid

Figure 5.1: A pictorial view of sampling from a RBM during RBM learning (cour-
tesy of Geoff Hinton).

Here, (v1 , h1 ) is a sample from the model, as a very rough estimate

of Emodel (vi hj ). The use of (v1 , h1 ) to approximate Emodel (vi hj ) gives
rise to the algorithm of CD-1. The sampling process can be pictorially
depicted in Figure 5.1.
Note that CD-k generalizes this to more steps of the Markov chain.
There are other techniques for estimating the log-likelihood gradient of
RBMs, in particular the stochastic maximum likelihood or persistent
contrastive divergence (PCD) [363, 406]. Both work better than CD
when using the RBM as a generative model.
Careful training of RBMs is essential to the success of applying
RBM and related deep learning techniques to solve practical problems.
See Technical Report [159] for a very useful practical guide for training
RBMs.
The RBM discussed above is both a generative and an unsupervised
model, which characterizes the input data distribution using hidden
variables and there is no label information involved. However, when
the label information is available, it can be used together with the
data to form the concatenated “data” set. Then the same CD learn-
ing can be applied to optimize the approximate “generative” objective
function related to data likelihood. Further, and more interestingly, a
“discriminative” objective function can be defined in terms of condi-
tional likelihood of labels. This discriminative RBM can be used to
“fine tune” RBM for classification tasks [203].
Ranzato et al. [297, 295] proposed an unsupervised learning algo-
rithm called sparse encoding symmetric machine (SESM), which is
quite similar to RBM. They both have a symmetric encoder and
5.2. Unsupervised layer-wise pre-training 245

decoder, and a logistic nonlinearity on the top of the encoder. The main
difference is that whereas the RBM is trained using (very approximate)
maximum likelihood, SESM is trained by simply minimizing the aver-
age energy plus an additional code sparsity term. SESM relies on the
sparsity term to prevent flat energy surfaces, while RBM relies on an
explicit contrastive term in the loss, an approximation of the log par-
tition function. Another difference is in the coding strategy in that the
code units are “noisy” and binary in the RBM, while they are quasi-
binary and sparse in SESM. The use of SESM in pre-training DNNs
for speech recognition can be found in [284].

5.2 Unsupervised layer-wise pre-training

Here we describe how to stack up RBMs just described to form a

DBN as the basis for DNN’s pre-training. Before delving into details,
we first note that this procedure, proposed by Hinton and Salakhut-
dinov [163] is a more general technique of unsupervised layer-wise
pretraining. That is, not only RBMs can be stacked to form deep gen-
erative (or discriminative) networks, but other types of networks can
also do the same, such as autoencoder variants as proposed by Bengio
et al. [28].
Stacking a number of the RBMs learned layer by layer from bottom
up gives rise to a DBN, an example of which is shown in Figure 5.2. The
stacking procedure is as follows. After learning a Gaussian-Bernoulli
RBM (for applications with continuous features such as speech) or
Bernoulli-Bernoulli RBM (for applications with nominal or binary fea-
tures such as black-white image or coded text), we treat the activation
probabilities of its hidden units as the data for training the Bernoulli-
Bernoulli RBM one layer up. The activation probabilities of the second-
layer Bernoulli-Bernoulli RBM are then used as the visible data input
for the third-layer Bernoulli-Bernoulli RBM, and so on. Some theoret-
ical justification of this efficient layer-by-layer greedy learning strat-
egy is given in [163], where it is shown that the stacking procedure
above improves a variational lower bound on the likelihood of the train-
ing data under the composite model. That is, the greedy procedure
246 Pre-Trained Deep Neural Networks — A Hybrid

Figure 5.2: An illustration of the DBN-DNN architecture.

above achieves approximate maximum likelihood learning. Note that

this learning procedure is unsupervised and requires no class label.
When applied to classification tasks, the generative pre-training
can be followed by or combined with other, typically discriminative,
learning procedures that fine-tune all of the weights jointly to improve
the performance of the network. This discriminative fine-tuning is per-
formed by adding a final layer of variables that represent the desired
outputs or labels provided in the training data. Then, the back-
propagation algorithm can be used to adjust or fine-tune the network
weights in the same way as for the standard feed-forward neural net-
work. What goes to the top, label layer of this DNN depends on the
application. For speech recognition applications, the top layer, denoted
by “l1 , l2 , . . . , lj , . . . , lL ,” in Figure 5.2, can represent either syllables,
phones, sub-phones, phone states, or other speech units used in the
HMM-based speech recognition system.
The generative pre-training described above has produced better
phone and speech recognition results than random initialization on
a wide variety of tasks, which will be surveyed in Section 7. Fur-
ther research has also shown the effectiveness of other pre-training
strategies. As an example, greedy layer-by-layer training may be carried
5.2. Unsupervised layer-wise pre-training 247

out with an additional discriminative term to the generative cost func-

tion at each level. And without generative pre-training, purely discrim-
inative training of DNNs from random initial weights using the tradi-
tional stochastic gradient decent method has been shown to work very
well when the scales of the initial weights are set carefully and the mini-
batch sizes, which trade off noisy gradients with convergence speed,
used in stochastic gradient decent are adapted prudently (e.g., with
an increasing size over training epochs). Also, randomization order in
creating mini-batches needs to be judiciously determined. Importantly,
it was found effective to learn a DNN by starting with a shallow neural
network with a single hidden layer. Once this has been trained discrimi-
natively (using early stops to avoid overfitting), a second hidden layer is
inserted between the first hidden layer and the labeled softmax output
units and the expanded deeper network is again trained discrimina-
tively. This can be continued until the desired number of hidden layers
is reached, after which a full backpropagation “fine tuning” is applied.
This discriminative “pre-training” procedure is found to work well in
practice [324, 419], especially with a reasonably large amount of train-
ing data. When the amount of training data is increased even more,
then some carefully designed random initialization methods can work
well also without using the above pre-training schemes.
In any case, pre-training based on the use of RBMs to stack up in
forming the DBN has been found to work well in most cases, regardless
of a large or small amount of training data. It is useful to point out
that there are other ways to perform pre-training in addition to the
use of RBMs and DBNs. For example, denoising autoencoders have
now been shown to be consistent estimators of the data generating
distribution [30]. Like RBMs, they are also shown to be generative
models from which one can sample. Unlike RBMs, however, an
unbiased estimator of the gradient of the training objective function
can be obtained by the denoising autoencoders, avoiding the need for
MCMC or variational approximations in the inner loop of training.
Therefore, the greedy layer-wise pre-training may be performed as
effectively by stacking the denoising autoencoders as by stacking the
RBMs each as a single-layer learner.
248 Pre-Trained Deep Neural Networks — A Hybrid

Further, a general framework for layer-wise pre-training can be

found in many deep learning papers; e.g., Section 2 of [21]. This
includes, as a special case, the use of RBMs as the single-layer build-
ing block as discussed in this section. The more general framework
can cover the RBM/DBN as well as any other unsupervised feature
extractor. It can also cover the case of unsupervised pre-training of the
representation only followed by a separate stage of learning a classifier
on top of the unsupervised, pre-trained features [215, 216, 217].

5.3 Interfacing DNNs with HMMs

The pre-trained DNN as a prominent example of the hybrid deep

networks discussed so far in this chapter is a static classifier with
input vectors having a fixed dimensionality. However, many practi-
cal pattern recognition and information processing problems, including
speech recognition, machine translation, natural language understand-
ing, video processing and bio-information processing, require sequence
recognition. In sequence recognition, sometimes called classification
with structured input/output, the dimensionality of both inputs and
outputs are variable.
The HMM, based on dynamic programing operations, is a con-
venient tool to help port the strength of a static classifier to han-
dle dynamic or sequential patterns. Thus, it is natural to combine
feed-forward neural networks and HMMs to bridge the gap between
the static and sequence pattern recognition, as was done in the early
days of neural networks for speech recognition [17, 25, 42]. A popu-
lar architecture to fulfill this role with the use of the DNN is shown
in Figure 5.3. This architecture has been successfully used in speech
recognition experiments as reported in [67, 68].
It is important to note that the unique elasticity of temporal dynam-
ics of speech as elaborated in [45, 73, 76, 83] would require temporally
5.3. Interfacing DNNs with HMMs 249

Figure 5.3: Interface between DBN/DNN and HMM to form a DNN–HMM. This
architecture, developed at Microsoft, has been successfully used in speech recognition
experiments reported in [67, 68]. [after [67, 68], @IEEE].

correlated models more powerful than HMMs for the ultimate success
of speech recognition. Integrating such dynamic models that have real-
istic co-articulatory properties with the DNN and possibly other deep
learning models to form the coherent dynamic deep architecture is a
challenging new research direction.
 6
Deep Stacking Networks and Variants —
Supervised Learning

6.1 Introduction

While the DNN just reviewed has been shown to be extremely power-
ful in connection with performing recognition and classification tasks
including speech recognition and image classification, training a DNN
has proven to be difficult computationally. In particular, conventional
techniques for training DNNs at the fine tuning phase involve the uti-
lization of a stochastic gradient descent learning algorithm, which is
difficult to parallelize across machines. This makes learning at large
scale nontrivial. For example, it has been possible to use one single,
very powerful GPU machine to train DNN-based speech recognizers
with dozens to a few hundreds or thousands of hours of speech training
data with remarkable results. It is less clear, however, how to scale up
this success with much more training data. See [69] for recent work in
this direction.
Here we describe a new deep learning architecture, the deep stacking
network (DSN), which was originally designed with the learning scal-
ability problem in mind. This chapter is based in part on the recent
publications of [106, 110, 180, 181] with expanded discussions.

250
6.1. Introduction 251

The central idea of the DSN design relates to the concept of stack-
ing, as proposed and explored in [28, 44, 392], where simple modules of
functions or classifiers are composed first and then they are “stacked”
on top of each other in order to learn complex functions or classifiers.
Various ways of implementing stacking operations have been developed
in the past, typically making use of supervised information in the sim-
ple modules. The new features for the stacked classifier at a higher
level of the stacking architecture often come from concatenation of the
classifier output of a lower module and the raw input features. In [60],
the simple module used for stacking was a conditional random field
(CRF). This type of deep architecture was further developed with hid-
den states added for successful natural language and speech recognition
applications where segmentation information is unknown in the train-
ing data [429]. Convolutional neural networks, as in [185], can also be
considered as a stacking architecture but the supervision information
is typically not used until in the final stacking module.
The DSN architecture was originally presented in [106] and was
referred as deep convex network or DCN to emphasize the convex
nature of a major portion of the algorithm used for learning the net-
work. The DSN makes use of supervision information for stacking each
of the basic modules, which takes the simplified form of multilayer per-
ceptron. In the basic module, the output units are linear and the hidden
units are sigmoidal nonlinear. The linearity in the output units permits
highly efficient, parallelizable, and closed-form estimation (a result of
convex optimization) for the output network weights given the hidden
units’ activities. Due to the closed-form constraints between the input
and output weights, the input weights can also be elegantly estimated in
an efficient, parallelizable, batch-mode manner, which we will describe
in some detail in Section 6.3.
The name “convex” used in [106] accentuates the role of convex
optimization in learning the output network weights given the hidden
units’ activities in each basic module. It also points to the importance
of the closed-form constraints, derived from the convexity, between the
input and output weights. Such constraints make the learning of the
remaining network parameters (i.e., the input network weights) much
easier than otherwise, enabling batch-mode learning of the DSN that
252 Deep Stacking Networks and Variants — Supervised Learning

can be distributed over CPU clusters. And in more recent publications,

the DSN was used when the key operation of stacking is emphasized.

6.2 A basic architecture of the deep stacking network

A DSN, as shown in Figure 6.1, includes a variable number of layered

modules, wherein each module is a specialized neural network con-
sisting of a single hidden layer and two trainable sets of weights. In
Figure 6.1, only four such modules are illustrated, where each module
is shown with a separate color. In practice, up to a few hundreds of
modules have been efficiently trained and used in image and speech
classification experiments.
The lowest module in the DSN comprises a linear layer with a set of
linear input units, a hidden nonlinear layer with a set of nonlinear units,
and a second linear layer with a set of linear output units. A sigmoidal
nonlinearity is typically used in the hidden layer. However, other non-
linearities can also be used. If the DSN is utilized in connection with rec-
ognizing an image, the input units can correspond to a number of pixels
(or extracted features) in the image, and can be assigned values based at
least in part upon intensity values, RGB values, or the like correspond-
ing to the respective pixels. If the DSN is utilized in connection with
speech recognition, the set of input units may correspond to samples
of speech waveform, or the extracted features from speech waveforms,
such as power spectra or cepstral coefficients. The output units in the
linear output layer represent the targets of classification. For instance,
if the DSN is configured to perform digit recognition, then the output
units may be representative of the values 0, 1, 2, 3, and so forth up to 9
with a 0–1 coding scheme. If the DSN is configured to perform speech
recognition, then the output units may be representative of phones,
HMM states of phones, or context-dependent HMM states of phones.
The lower-layer weight matrix, which we denote by W , connects
the linear input layer and the hidden nonlinear layer. The upper-layer
weight matrix, which we denote by U , connects the nonlinear hid-
den layer with the linear output layer. The weight matrix U can be
determined through a closed-form solution given the weight matrix W
when the mean square error training criterion is used.
6.2. A basic architecture of the deep stacking network 253

Figure 6.1: A DSN architecture using input–output stacking. Four modules are
illustrated, each with a distinct color. Dashed lines denote copying layers. [after
[366], @IEEE].

As indicated above, the DSN includes a set of serially connected,

overlapping, and layered modules, wherein each module has the same
architecture — a linear input layer followed by a nonlinear hidden
layer, which is connected to a linear output layer. Note that the output
units of a lower module are a subset of the input units of an adjacent
higher module in the DSN. More specifically, in a second module that
is directly above the lowest module in the DSN, the input units can
include the output units of the lowest module and optionally the raw
input feature.
This pattern of including output units in a lower module as a
portion of the input units in an adjacent higher module and thereafter
254 Deep Stacking Networks and Variants — Supervised Learning

learning a weight matrix that describes connection weights between

hidden units and linear output units via convex optimization can con-
tinue for many modules. A resultant learned DSN may then be deployed
in connection with an automatic classification task such as frame-level
speech phone or state classification. Connecting the DSN’s output to an
HMM or any dynamic programming device enables continuous speech
recognition and other forms of sequential pattern recognition.

6.3 A method for learning the DSN weights

Here, we provide some technical details on how the use of linear out-
put units in the DSN facilitates the learning of the DSN weights. A
single module is used to illustrate the advantage for simplicity rea-
sons. First, it is clear that the upper layer weight matrix U can
be efficiently learned once the activity matrix H over all training
samples in the hidden layer is known. Let’s denote the training vec-
tors by X = [x 1 , . . . , x i , . . . , x N ], in which each vector is denoted by
x i = [x1i , . . . , xji , . . . , xDi ]T where D is the dimension of the input vec-
tor, which is a function of the block, and N is the total number of
training samples. Denote by L the number of hidden units and by C
the dimension of the output vector. Then the output of a DSN block is
y i = U T h i where h i = σ(W T x i ) is the hidden-layer vector for sample
i, U is an L × C weight matrix at the upper layer of a block. W is a
D × L weight matrix at the lower layer of a block, and σ(·) is a sigmoid
function. Bias terms are implicitly represented in the above formulation
if x i and h i are augmented with ones.
Given target vectors in the full training set with a total of
N samples, T = [t 1 , . . . , t i , . . . , t N ], where each vector is t i =
[t1i , · · · , tji , . . . , tCi ]T , the parameters U and W are learned so as to
minimize the average of the total square error below:
1 1
E= y i − t i 2 = Tr[(Y − T )(Y − T )T ]
2 i 2

where the output of the network is

y i = U T h i = U T σ(W T x i ) = Gi (UW )
6.4. The tensor deep stacking network 255

which depends on both weight matrices, as in the standard neural net.

Assuming H = [h 1 , . . . , h i , . . . , h N ] is known, or equivalently, W is
known. Then, setting the error derivative with respective to U to zero
gives

U = (HH T )−1 HT T = F(W ), where h i = σ(W T x i ).

This provides an explicit constraint between U and W which were

treated independently in the conventional backpropagation algorithm.
Now, given the equality constraint U = F (W ), let’s use Lagrangian
multiplier method to solve the optimization problem in learning W
Optimizing the Lagrangian:
1
E= Gi (U , W ) − t i 2 + λU − F(W )
2 i

we can derive batch-mode gradient descent learning algorithm where

the gradient takes the following form [106, 413]:
∂E
= 2X[H T ◦ (1 − H )T ◦ [H † (HT T )(TH † ) − T T (TH † )]],
∂W
where H † = H T (H H T )−1 is pseudo-inverse of H and symbol ◦
denotes element-wise multiplication.
Compared with conventional backpropagation, the above method
has less noise in gradient computation due to the exploitation of the
explicit constraint U = F (W ). As such, it was found experimentally
that, unlike backpropagation, batch training is effective, which aids
parallel learning of the DSN.

6.4 The tensor deep stacking network

The above DSN architecture has recently been generalized to its ten-
sorized version, which we call the tensor DSN (TDSN) [180, 181]. It
has the same scalability as the DSN in terms of parallelizability in
learning, but it generalizes the DSN by providing higher-order feature
interactions missing in the DSN.
The architecture of the TDSN is similar to that of the DSN in the
way that stacking operation is carried out. That is, modules of the
256 Deep Stacking Networks and Variants — Supervised Learning

Figure 6.2: Comparisons of a single module of a DSN (left) and that of a tensor
DSN (TDSN). Two equivalent forms of a TDSN module are shown to the right.
[after [180], @IEEE].

TDSN are stacked up in a similar way to form a deep architecture.

The differences between the TDSN and the DSN lie mainly in how
each module is constructed. In the DSN, we have one set of hidden
units forming a hidden layer, as denoted at the left panel of Figure 6.2.
In contrast, each module of a TDSN contains two independent hidden
layers, denoted as “Hidden 1” and “Hidden 2” in the middle and right
panels of Figure 6.2. As a result of this difference, the upper-layer
weights, denoted by “U” in Figure 6.2, changes from a matrix (a two
dimensional array) in the DSN to a tensor (a three dimensional array)
in the TDSN, shown as a cube labeled by “U” in the middle panel.
The tensor U has a three-way connection, one to the prediction
layer and the remaining to the two separate hidden layers. An equiva-
lent form of this TDSN module is shown in the right panel of Figure 6.2,
where the implicit hidden layer is formed by expanding the two sepa-
rate hidden layers into their outer product. The resulting large vector
contains all possible pair-wise products for the two sets of hidden-layer
vectors. This turns tensor U into a matrix again whose dimensions are
(1) size of the prediction layer; and (2) product of the two hidden lay-
ers’ sizes. Such equivalence enables the same convex optimization for
learning U developed for the DSN to be applied to learning tensor U.
Importantly, higher-order hidden feature interactions are enabled in
the TDSN via the outer product construction for the large, implicit
hidden layer.
6.5. The Kernelized deep stacking network 257

Figure 6.3: Stacking of TDSN modules by concatenating prediction vector with

input vector. [after [180], @IEEE].

Stacking the TDSN modules to form a deep architecture pursues in

a similar way to the DSN by concatenating various vectors. Two exam-
ples are shown in Figures 6.3 and 6.4. Note stacking by concatenating
hidden layers with input (Figure 6.4) would be difficult for the DSN
since its hidden layer tends to be too large for practical purposes.

6.5 The Kernelized deep stacking network

The DSN architecture has also recently been generalized to its ker-
nelized version, which we call the kernel-DSN (K-DSN) [102, 171]. The
motivation of the extension is to increase the size of the hidden units in
each DSN module, yet without increasing the size of the free parameters
to learn. This goal can be easily accomplished using the kernel trick,
resulting in the K-DSN which we describe below.
258 Deep Stacking Networks and Variants — Supervised Learning

Figure 6.4: Stacking of TDSN modules by concatenating two hidden-layers’ vectors

with the input vector.

In the DSN architecture reviewed above optimizing the weight

matrix U given the hidden layers’ outputs in each module is a con-
vex optimization problem. However, the problem of optimizing weight
matrix W and thus the whole network is nonconvex. In a recent exten-
sion of DSN, a tensor structure was imposed, shifting most of the
nonconvex learning burden for W to the convex optimization of U
[180, 181]. In the new K-DSN extension, we completely eliminate non-
convex learning for W using the kernel trick.
To derive the K-DSN architecture and the associated learning algo-
rithm, we first take the bottom module of DSN as an example and
generalize the sigmoidal hidden layer hi = σ(W T xi ) in the DSN mod-
ule into a generic nonlinear mapping function G(X) from the raw
input feature X, with high dimensionality in G(X) (possibly infinite)
determined only implicitly by a kernel function to be chosen. Second,
6.5. The Kernelized deep stacking network 259

we formulate the constrained optimization problem of

1 C
minimize Tr[EE T ] + U T U
2 2
subject to T − U G(X) = E.
T

Third, we make use of dual representations of the above constrained

optimization problem to obtain U = GT a, where vector a takes the
following form:
a = (CI + K )−1 T
and K = G(X )G T (X) is a symmetric kernel matrix with elements
Knm = gT (xn )g(xm ).
Finally, for each new input vector x in the test or dev set, we obtain
the K-DSN (bottom) module’s prediction as
y(x) = U T g(x) = aT G(X)g(x) = k T (x)(CI + K )−1 T ,
where the kernel vector k(x) is so defined that its elements have values
of kn (x) = k(x n , x) in which x n is a training sample and x is the
current test sample.
For lth module in K-DCN where l ≥ 2, the kernel matrix is
modified to
K = G([X |Y (l−1) | Y (l−2) | . . . Y (1) ])G T ([X |Y (l−1) |Y (l−2) | . . . Y (1) ]).
The key advantages of K-DSN can be analyzed as follows. First,
unlike DSN which needs to compute hidden units’ output, the K-DSN
does not need to explicitly compute hidden units’ output G(X) or
G([X |Y (l−1) |Y (l−2) | . . . Y (1) ]). When Gaussian kernels are used, ker-
nel trick equivalently gives us an infinite number of hidden units with-
out the need to compute them explicitly. Further, we no longer need
to learn the lower-layer weight matrix W in DSN as described in [102]
and the kernel parameter (e.g., the single variance parameter σ in the
Gaussian kernel) makes K-DSN much less subject to overfitting than
DSN. Figure 6.5 illustrates the basic architecture of a K-DSN using the
Gaussian kernel and using three modules.
The entire K-DSN with Gaussian kernels is characterized by two
sets of module-dependent hyper-parameters: σ (l) and C (l) the kernel
260 Deep Stacking Networks and Variants — Supervised Learning

Figure 6.5: An example architecture of the K-DSN with three modules each of
which uses a Gaussian kernel with different kernel parameters. [after [102], @IEEE].

smoothing parameter and regularization parameter, respectively. While

both parameters are intuitive and their tuning (via line search or leave-
one-out cross validation) is straightforward for a single bottom mod-
ule, tuning the full network with all the modules is more difficult. For
example, if the bottom module is tuned too well, then adding more
modules would not benefit much. In contrast, when the lower modules
are loosely tuned (i.e., relaxed from the results obtained from straight-
forward methods), the overall K-DSN often performs much better. The
experimental results reported by Deng et al. [102] are obtained using a
set of empirically determined tuning schedules to adaptively regularize
the K-DSN from bottom to top modules.
The K-DSN described here has a set of highly desirable proper-
ties from the machine learning and pattern recognition perspectives. It
combines the power of deep learning and kernel learning in a principled
6.5. The Kernelized deep stacking network 261

way and unlike the basic DSN there is no longer nonconvex optimiza-
tion problem involved in training the K-DSN. The computation steps
make the K-DSN easier to scale up for parallel computing in distributed
servers than the DSN and tensor-DSN. There are many fewer param-
eters in the K-DSN to tune than in the DSN, T-DSN, and DNN, and
there is no need for pre-training. It is found in the study of [102] that
regularization plays a much more important role in the K-DSN than
in the basic DSN and Tensor-DSN. Further, effective regularization
schedules developed for learning the K-DSN weights can be motivated
by intuitive insight from useful optimization tricks such as the heuristic
in Rprop or resilient backpropagation algorithm [302].
However, as inherent in any kernel method, the scalability becomes
an issue also for the K-DSN as the training and testing samples become
very large. A solution is provided in the study by Huang et al. [171],
based on the use of random Fourier features, which possess the strong
theoretical property of approximating the Gaussian kernel while render-
ing efficient computation in both training and evaluation of the K-DSN
with large training samples. It is empirically demonstrated that just like
the conventional K-DSN exploiting rigorous Gaussian kernels, the use
of random Fourier features also enables successful stacking of kernel
modules to form a deep architecture.
 7
Selected Applications in Speech
and Audio Processing

7.1 Acoustic modeling for speech recognition

As discussed in Section 2, speech recognition is the very first success-

ful application of deep learning methods at an industry scale. This
success is a result of close academic-industrial collaboration, initiated
at Microsoft Research, with the involved researchers identifying and
acutely attending to the industrial need for large-scale deployment
[68, 89, 109, 161, 323, 414]. It is also a result of carefully exploiting
the strengths of the deep learning and the then-state-of-the-art speech
recognition technology, including notably the highly efficient decoding
techniques.
Speech recognition has long been dominated by the GMM–HMM
method, with an underlying shallow or flat generative model of context-
dependent GMMs and HMMs (e.g., [92, 93, 187, 293]). Neural networks
once were a popular approach but had not been competitive with the
GMM–HMM [42, 87, 261, 382]. Generative models with deep hidden
dynamics likewise have also not been clearly competitive (e.g., [45, 73,
108, 282]).
Deep learning and the DNN started making their impact in speech
recognition in 2010, after close collaborations between academic and

262
7.1. Acoustic modeling for speech recognition 263

industrial researchers; see reviews in [89, 161]. The collaborative work

started in phone recognition tasks [89, 100, 135, 136, 257, 260, 258,
309, 311, 334], demonstrating the power of hybrid DNN architec-
tures discussed in Section 5 and of subsequent new architectures
with convolutional and recurrent structure. The work also showed
the importance of raw speech features of spectrogram — back from
the long-popular MFCC features toward but not yet reaching the
raw speech-waveform level [183, 327]. The collaboration continued to
large vocabulary tasks with more convincing, highly positive results
[67, 68, 94, 89, 161, 199, 195, 223, 323, 353, 399, 414]. The success in
large vocabulary speech recognition is in large part attributed to the
use of a very large DNN output layer structured in the same way as
the GMM–HMM speech units (senones), motivated partially by the
speech researchers’ desires to take advantage of the context-dependent
phone modeling techniques that have been proven to work well in the
GMM–HMM framework, and to keep the change of the already highly
efficient decoder software’s infrastructure developed for the GMM–
HMM systems to a minimum. In the meantime, this body of work
also demonstrated the possibility to reduce the need for the DBN-
like pre-training in effective learning of DNNs when a large amount
of labeled data is available. A combination of three factors helped
to quickly spread the success of deep learning in speech recognition
to the entire speech industry and academia: (1) significantly lowered
errors compared with the then-state-of-the-art GMM-HMM systems;
(2) minimal decoder changes required to deploy the new DNN-based
speech recognizer due to the use of senones as the DNN output; and
(3) reduced system complexity empowered by the DNN’s strong mod-
eling power. By the ICASSP-2013 timeframe, at least 15 major speech
recognition groups worldwide confirmed experimentally the success of
DNNs with very large tasks and with the use of raw speech spectral
features other than MFCCs. The most notable groups include major
industrial speech labs worldwide: Microsoft [49, 89, 94, 324, 399, 430],
IBM [195, 309, 311, 307, 317], Google [69, 150, 184, 223], iFlyTek, and
Baidu. Their results represent a new state-of-the-art in speech recog-
nition widely deployed in these companies’ voice products and services
with extensive media coverage in recent years.
264 Selected Applications in Speech and Audio Processing

In the remainder of this chapter, we review a wide range of speech

recognition work based on deep learning methods according to several
major themes expressed in the section titles.

7.1.1 Back to primitive spectral features of speech

Deep learning, also referred as representation learning or (unsuper-

vised) feature learning, sets an important goal of automatic discovery
of powerful features from raw input data independent of application
domains. For speech feature learning and for speech recognition, this
goal is condensed to the use of primitive spectral or possibly wave-
form features. Over the past 30 years or so, largely “hand-crafted”
transformations of speech spectrogram have led to significant accuracy
improvements in the GMM-based HMM systems, despite the known
loss of information from the raw speech data. The most successful
transformation is the non-adaptive cosine transform, which gave rise to
Mel-frequency cepstral coefficients (MFCC) features. The cosine trans-
form approximately de-correlates feature components, which is impor-
tant for the use of GMMs with diagonal covariance matrices. However,
when GMMs are replaced by deep learning models such as DNNs, deep
belief nets (DBNs), or deep autoencoders, such de-correlation becomes
irrelevant due to the very strength of the deep learning methods in
modeling data correlation. As discussed in detail in Section 4, early
work of [100] demonstrated this strength and in particular the benefit
of spectrograms over MFCCs in effective coding of bottleneck speech
features using autoencoders in an unsupervised manner.
The pipeline from speech waveforms (raw speech features) to
MFCCs and their temporal differences goes through intermediate
stages of log-spectra and then (Mel-warped) filter-banks, with learned
parameters based on the data. An important character of deep learn-
ing is to move away from separate design of feature representations and
of classifiers. This idea of jointly learning classifier and feature trans-
formation for speech recognition was already explored in early studies
on the GMM–HMM based systems; e.g., [33, 50, 51, 299]. However,
greater speech recognition performance gain is obtained only recently
7.1. Acoustic modeling for speech recognition 265

in the recognizers empowered by deep learning methods. For example,

Mohamed et al. [259], Li et al. [221], and Deng et al. [94] showed signif-
icantly lowered speech recognition errors using large-scale DNNs when
moving from the MFCC features back to more primitive (Mel-scaled)
filter-bank features. These results indicate that DNNs can learn a bet-
ter transformation than the original fixed cosine transform from the
Mel-scaled filter-bank features.
Compared with MFCCs, “raw” spectral features not only retain
more information, but also enable the use of convolution and pool-
ing operations to represent and handle some typical speech variabil-
ity — e.g., vocal tract length differences across speakers, distinct speak-
ing styles causing formant undershoot or overshoot, etc. — expressed
explicitly in the frequency domain. For example, the convolutional neu-
ral network (CNN) can only be meaningfully and effectively applied to
speech recognition [1, 2, 3, 94] when spectral features, instead of MFCC
features, are used.
More recently, Sainath et al. [307] went one step further toward
raw features by learning the parameters that define the filter-banks
on power spectra. That is, rather than using Mel-warped filter-bank
features as the input features as in [1, 3, 50, 221], the weights corre-
sponding to the Mel-scale filters are only used to initialize the param-
eters, which are subsequently learned together with the rest of the
deep network as the classifier. The overall architecture of the jointly
learned feature generator and classifier is shown in Figure 7.1. Substan-
tial speech recognition error reduction is reported in [307].
It has been shown that not only learning the spectral aspect of
the features are beneficial for speech recognition, learning the tempo-
ral aspect of the features is also helpful [332]. Further, Yu et al. [426]
carefully analyzed the properties of different layers in the DNN as the
layer-wise extracted features starting from the lower raw filter-bank
features. They found that the improved speech recognition accuracy
achieved by the DNNs partially attributes to DNN’s ability to extract
discriminative internal representations that are robust to the many
sources of variability in speech signals. They also show that these rep-
resentations become increasingly insensitive to small perturbations in
266 Selected Applications in Speech and Audio Processing

Figure 7.1: Illustration of the joint learning of filter parameters and the rest of the
deep network. [after [307], @IEEE].

the input at higher layers, which helps to achieve better speech recog-
nition accuracy.
To the extreme end, deep learning would promote to use the lowest
level of raw features of speech, i.e., speech sound waveforms, for speech
recognition, and learn the transformation automatically. As an initial
attempt toward this goal the study carried out by Jaitly and Hinton
[183] makes use of speech sound waves as the raw input feature to an
RBM with a convolutional structure as the classifier. With the use
of rectified linear units in the hidden layer [130], it is possible, to a
limited extent, to automatically normalize the amplitude variation
in the waveform signal. Although the final results are disappointing,
the work shows that much work is needed along this direction. For
example, just as demonstrated by Sainath et al. [307] that the use of
raw spectra as features requires additional attention in normalization
than MFCCs, the use of speech waveforms demands even more
attention in normalization [327]. This is true for both GMM-based
and deep learning based methods.
7.1. Acoustic modeling for speech recognition 267

7.1.2 The DNN–HMM architecture versus use of

DNN-derived features

Another major theme in the recent studies reported in the literature on

applying deep learning methods to speech recognition is two disparate
ways of using the DNN: (1) Direct applications of the DNN-HMM
architecture as discussed in Section 5.3 to perform speech recognition;
and (2) The use of DNNs to extract or derive features, which are then
fed into a separate sequence classifier. In the speech recognition lit-
erature [42], a system, in which a neural network’s output is directly
used to estimate the emission probabilities of an HMM, is often called
an ANN/HMM hybrid system. This should be distinguished from the
use of “hybrid” in Section 5 and throughout this monograph, where
a hybrid of unsupervised pre-training and of supervised fine tuning is
exploited to learn the parameters of DNNs.

7.1.2.1 The DNN–HMM architecture as a recognizer

An early DNN–HMM architecture [257] was presented at the NIPS

Workshop [109], developed, analyzed, and assisted by University of
Toronto and MSR speech researchers. In this work, a five-layer DNN
(called the DBN in the paper) was used to replace the Gaussian mixture
models in the GMM–HMM system, and the monophone state was used
as the modeling unit. Although monophones are generally accepted
as a weaker phonetic representation than triphones, the DNN–HMM
approach with monophones was shown to achieve higher phone recog-
nition accuracy than the state-of-the-art triphone GMM–HMM sys-
tems. Further, the DNN results were found to be slightly superior to
the then-best-performing single system based on the generative hid-
den trajectory model (HTM) in the literature [105, 108] evaluated on
the same, commonly used TIMIT task by many speech researchers
[107, 108, 274, 313]. At MSR, Redmond, the error patterns produced
by these two separate systems (the DNN vs. the HTM) were carefully
analyzed and found to be very different, reflecting distinct core capa-
bilities of the two approaches and igniting intensive further studies on
the DNN–HMM approach described below.
268 Selected Applications in Speech and Audio Processing

MSR and University of Toronto researchers [67, 68, 414] extended

the DNN–HMM system from the monophone phonetic representation of
the DNN outputs to the triphone or context-dependent counterpart and
from phone recognition to large vocabulary speech recognition. Experi-
ments conducted at MSR on the 24-hour and 48-hour Bing mobile voice
search datasets collected under the real usage scenario demonstrate
that the context-dependent DNN–HMM significantly outperforms the
state-of-the-art GMM-HMM system. Three factors, in addition to the
use of the DNN, contribute to the success: the use of tied triphones
as the DNN modeling units, the use of the best available tri-phone
GMM–HMM to generate the tri-phone state alignment, and the effec-
tive exploitation of a long window of input features. Experiments also
indicate that the decoding time of a five-layer DNN–HMM is almost
the same as that of the state-of-the-art triphone GMM–HMM.
The success was quickly extended to large vocabulary speech recog-
nition tasks with hundreds and even thousands of hours of training set
and with thousands of tri-phone states, including the Switchboard and
Broadcast News databases, and Google’s voice search and YouTube
tasks [94, 161, 184, 309, 311, 324]. For example, on the Switchboard
benchmark, the context-dependent DNN–HMM (CD-DNN–HMM) is
shown to cut error by one third compared to the state-of-the-art GMM–
HMM system [323]. As a summary, we show in Table 7.1 some quanti-
tative recognition error rates in relatively early literature produced by
the basic DNN–HMM architecture in comparison with those by the pre-
vious state-of-the-art systems based on the generative models. (More
advanced architectures have produced better results than shown here).
Note from sub-tables A to D, the training data are increased approx-
imately one order of magnitude from one task to the next. Not only
the computation scales up well (i.e., almost linearly) with the training
size, but most importantly the relative error rate reduction increases
substantially with increasing amounts of training data — from approx-
imately 10% to 20%, and then to 30%. This set of results highlight the
strongly desirable properties of the DNN-based methods, despite the
conceptual simplicity of the overall DNN–HMM architecture and some
known weaknesses.
7.1. Acoustic modeling for speech recognition 269

Table 7.1: Comparisons of the DNN–HMM architecture with the generative model
(e.g., the GMM–HMM) in terms of phone or word recognition error rates. From
sub-tables A to D, the training data are increased approximately three orders of
magnitudes.

Features Setup Error Rates

A: TIMIT Phone recognition (3 hours of training)
GMM w. Hidden dynamics 24.8%
DNN 5 layers × 2048 23.0%
B: Voice Search SER (24–48 hours of training)
GMM MPE (760 24-mix) 36.2%
DNN 5 layers × 2048 30.1%
C: Switch Board WER (309 hours of training)
GMM BMMI (9K 40-mix) 23.6%
DNN 7 layers × 2048 15.8%
D: Switch Board WER (2000 hours of training)
GMM BMMI (18K 72-mix) 21.7%
DNN 7 layers × 2048 14.6%

7.1.2.2 The use of DNN-derived features in a separate recognizer

One clear weakness of the above DNN–HMM architecture for speech

recognition is that much of the highly effective techniques for the
GMM–HMM systems, including discriminative training (in both fea-
ture space and model space), unsupervised speaker adaptation, noise
robustness, and scalable batch training tools for big training data,
developed over the past 20 some years may not be directly applica-
ble to the new systems although similar techniques have been recently
developed for DNN–HMMs. To remedy this problem, the “tandem”
approach, developed originally by Hermansky et al. [154], has been
adopted, where the output of the neural networks in the form of pos-
terior probabilities for phone classes, are used, often in conjunction
with the acoustic features to form new augmented input features, in a
separate GMM–HMM system.
270 Selected Applications in Speech and Audio Processing

This tandem approach is used by Vinyals and Ravuri [379] where a

DNN’s outputs are extracted to serve as the features for mismatched
noisy speech. It is reported that DNNs outperform the neural net-
works with a single hidden layer under the clean condition, but the
gains slowly diminish as the noise level is increased. Furthermore, using
MFCCs in conjunction with the posteriors computed from DNNs out-
performs using the DNN features alone in low to moderate noise con-
ditions with the tandem architecture. Comparisons of such tandem
approach with the direct DNN–HMM approach are made by Tüske
et al. [368] and Imseng et al. [182].
An alternative way of extracting the DNN features is to use the
“bottleneck” layer, which is narrower than other layers in the DNN,
to restrict the capacity of the network. Then, such bottleneck features
are fed to a GMM–HMM system, often in conjunction with the orig-
inal acoustic features and some dimensionality reduction techniques.
The bottleneck features derived from the DNN are believed to capture
information complementary to conventional acoustic features derived
from the short-time spectra of the input. A speech recognizer based on
the above bottleneck feature approach is built by Yu and Seltzer [425],
with the overall architecture shown in Figure 7.2. Several variants of
the DNN-based bottleneck-feature approach have been explored; see
details in [16, 137, 201, 285, 308, 368].
Yet another method to derive the features from the DNN is to feed
its top-most hidden layer as the new features for a separate speech

Figure 7.2: Illustration of the use of bottleneck (BN) features extracted from a
DNN in a GMM–HMM speech recognizer. [after [425], @IEEE].
7.1. Acoustic modeling for speech recognition 271

recognizer. In [399], a GMM–HMM is used as such a recognizer, and

the high-dimensional, DNN-derived features are subject to dimension-
ality reduction before feeding them into the recognizer. More recently,
a recurrent neural network (RNN) is used as the “backend” recognizer
receiving the high-dimensional, DNN-derived features as the input
without dimensionality reduction [48, 85]. These studies also show
that the use of the top-most hidden layer of the DNN as features is
better than other hidden layers and also better than the output layer
in terms of recognition accuracy for the RNN sequence classifier.

7.1.3 Noise robustness by deep learning

The study of noise robustness in speech recognition has a long his-

tory, mostly before the recent rise of deep learning. One major con-
tributing factor to the often observed brittleness of speech recogni-
tion technology is the inability of the standard GMM–HMM-based
acoustic model to accurately model noise-distorted speech test data
that differs in character from the training data, which may or may
not be distorted by noise. A wide range of noise-robust techniques
developed over past 30 years can be analyzed and categorized using
five different criteria: (1) feature-domain versus model-domain pro-
cessing, (2) the use of prior knowledge about the acoustic environ-
ment distortion, (3) the use of explicit environment-distortion mod-
els, (4) deterministic versus uncertainty processing, and (5) the use of
acoustic models trained jointly with the same feature enhancement or
model adaptation process used in the testing stage. See a comprehen-
sive review in [220] and some additional review literature or original
work in [4, 82, 119, 140, 230, 370, 404, 431, 444].
Many of the model-domain techniques developed for GMM–HMMs
(e.g., model-domain noise robustness techniques surveyed by Li et al.
[220] and Gales [119]) are not directly applicable to the new deep
learning models for speech recognition. The feature-domain techniques,
however, can be directly applied to the DNN system. A detailed inves-
tigation of the use of DNNs for noise robust speech recognition in the
feature domain was reported by Seltzer et al. [325], who applied the
C-MMSE [415] feature enhancement algorithm on the input feature
272 Selected Applications in Speech and Audio Processing

used in the DNN. By processing both the training and testing data
with the same algorithm, any consistent errors or artifacts introduced
by the enhancement algorithm can be learned by the DNN–HMM rec-
ognizer. This study also successfully explored the use of the noise aware
training paradigm for training the DNN, where each observation was
augmented with an estimate of the noise. Strong results were obtained
on the Aurora4 task. More recently, Kashiwagi et al. [191] applied the
SPLICE feature enhancement technique [82] to a DNN speech rec-
ognizer. In that study the DNN’s output layer was determined on
clean data instead of on noisy data as in the study reported by Seltzer
et al. [325].
Besides DNN, other deep architectures have also been proposed to
perform feature enhancement and noise-robust speech recognition. For
example, Mass et al. [235] applied a deep recurrent auto encoder neural
network to remove noise in the input features for robust speech recogni-
tion. The model was trained on stereo (noisy and clean) speech features
to predict clean features given noisy input, similar to the SPLICE setup
but using a deep model instead of a GMM. Vinyals and Ravuri [379]
investigated the tandem approaches to noise-robust speech recognition,
where DNNs were trained directly with noisy speech to generate pos-
terior features. Finally, Rennie et al. [300] explored the use of a version
of the RBM, called the factorial hidden RBM, for noise-robust speech
recognition.

7.1.4 Output representations in the DNN

Most deep learning methods for speech recognition and other infor-
mation processing applications have focused on learning represen-
tations from input acoustic features without paying attention to
output representations. The recent 2013 NIPS Workshop on Learning
Output Representations (http://nips.cc/Conferences/2013/Program/
event.php?ID=3714) was dedicated to bridging this gap. For exam-
ple, the Deep Visual-Semantic Embedding Model described in [117],
to be discussed more in Section 11) exploits continuous-valued out-
put representations obtained from the text embeddings to assist in the
7.1. Acoustic modeling for speech recognition 273

branch of the deep network for classifying images. For speech recogni-
tion, the importance of designing effective linguistic representations for
the output layers of deep networks is highlighted in [79].
Most current DNN systems use a high-dimensional output represen-
tation to match the context-dependent phonetic states in the HMMs.
For this reason, the output layer evaluation can cost 1/3 of the total
computation time. To improve the decoding speed, techniques such
as low-rank approximation is typically applied to the output layer.
In [310] and [397], the DNN with high-dimensional output layer was
trained first. The singular value decomposition (SVD)-based dimen-
sion reduction technique was then performed on the large output-layer
matrix. The resulting matrices are further combined and as the result
the original large weight matrix is approximated by a product of two
much smaller matrices. This technique in essence converts the origi-
nal large output layer to two layers — a bottleneck linear layer and
a nonlinear output layer — both with smaller weight matrices. The
converted DNN with reduced dimensionality is further refined. The
experimental results show that no speech recognition accuracy reduc-
tion was observed even when the size is cut to half, while the run-time
computation is significantly reduced.
The output representations for speech recognition can benefit from
the structured design of the symbolic or phonological units of speech
as presented in [79]. The rich phonological structure of symbolic nature
in human speech has been well known for many years. Likewise, it has
also been well understood for a long time that the use of phonetic
or its finer state sequences, even with contextual dependency, in engi-
neering speech recognition systems, is inadequate in representing such
rich structure [86, 273, 355], and thus leaving a promising open direc-
tion to improve the speech recognition systems’ performance. Basic
theories about the internal structure of speech sounds and their rel-
evance to speech recognition technology in terms of the specification,
design, and learning of possible output representations of the underly-
ing speech model for speech target sequences are surveyed in [76] and
more recently in [79].
There has been a growing body of deep learning work in speech
recognition with their focus placed on designing output representations
274 Selected Applications in Speech and Audio Processing

related to linguistic structure. In [383, 384], a limitation of the out-

put representation design, based on the context-dependent phone units
as proposed in [67, 68], is recognized and a solution is offered. The
root cause of this limitation is that all context-dependent phone states
within a cluster created by the decision tree share the same set of
parameters and this reduces its resolution power for fine-grained states
during the decoding phase. The solution proposed formulates output
representations of the context-dependent DNN as an instance of the
canonical state modeling technique, making use of broad phonetic
classes. First, triphones are clustered into multiple sets of shorter bi-
phones using broad phone contexts. Then, the DNN is trained to dis-
criminate the bi-phones within each set. Logistic regression is used
to transform the canonical states into the detailed triphone state
output probabilities. That is, the overall design of the output rep-
resentation of the context-dependent DNN is hierarchical in nature,
solving both the data sparseness and low-resolution problems at the
same time.
Related work on designing the output linguistic representations for
speech recognition can be found in [197] and in [241]. While the designs
are in the context of GMM–HMM-based speech recognition systems,
they both can be extended to deep learning models.

7.1.5 Adaptation of the DNN-based speech recognizers

The DNN–HMM is an advanced version of the artificial neural net-

work and HMM “hybrid” system developed in 1990s, for which several
adaptation techniques have been developed. Most of these techniques
are based on linear transformation of the network weights of either
input or output layers. A number of exploratory studies on DNN adap-
tation made use of the same or related linear transformation methods
[223, 401, 402]. However, compared with the earlier narrower and shal-
lower neural network systems, the DNN–HMM has significantly more
parameters due to wider and deeper hidden layers used and the much
larger output layer designed to model context dependent phones and
states. This difference casts special challenges to adapting the DNN–
HMM, especially when the adaptation data is small. Here we discuss
7.1. Acoustic modeling for speech recognition 275

representative recent studies on overcoming such challenges in adapting

the large-sized DNN weights in distinct ways.
Yu et al. [430] proposed a regularized adaptation technique for
DNNs. It adapts the DNN weights conservatively by forcing the distri-
bution estimated from the adapted model to be close to that estimated
from those before the adaptation. This constraint is realized by adding
Kullback–Leibler divergence (KLD) regularization to the adaptation
criterion. This type of regularization is shown to be equivalent to a
modification of the target distribution in the conventional backprop-
agation algorithm and thus the training of the DNN remains largely
unchanged. The new target distribution is derived to be a linear inter-
polation of the distribution estimated from the model before adaptation
and the ground truth alignment of the adaptation data. This interpola-
tion prevents overtraining by keeping the adapted model from straying
too far from the speaker-independent model. This type of adaptation
differs from L2 regularization, which constrains the model parameters
themselves rather than the output probabilities.
In [330], adaptation of the DNN was applied not on the conventional
network weights but on the hidden activation functions. In this way, the
main limitation of current adaptation techniques based on adaptable
linear transformation of the network weights in either the input or the
output layer is effectively overcome, since the new method only needs
to adapt a more limited number of hidden activation functions.
Several studies were carried out on unsupervised or semi-supervised
adaptation of DNN acoustic models with different types of input fea-
tures with success [223, 405].
Most recently, Saon et al. [317] explored a new and highly effective
method in adapting DNNs for speech recognition. The method com-
bined I-vector features with fMLLR (feature-domain max-likelihood
linear regression) features as the input into a DNN. I-vectors or
(speaker) identity vectors are commonly used for speaker verifica-
tion and speaker recognition applications, as they encapsulate relevant
information about a speaker’s identity in a low-dimensional feature
vector. The fMLLR is an effective adaptation technique developed for
GMM–HMM systems. Since I-vectors do not obey locality in frequency,
they must be combined carefully with the fMLLR features that obey
276 Selected Applications in Speech and Audio Processing

locality. The architecture of the multi-scale CNN–DNN was shown to

be effective for the combination of these two different types of features.
During both training and decoding, the speaker-specific I-vector was
appended to the frame-based fMLLR features.

7.1.6 Better architectures and nonlinear units

Over recent years, since the success of the (fully-connected) DNN–

HMM hybrid system was demonstrated in [67, 68, 109, 161, 257, 258,
308, 309, 324, 429], many new architectures and nonlinear units have
been proposed and evaluated for speech recognition. Here we provide
an overview of this progress, extending the overview provided in [89].
The tensor version of the DNN is reported by Yu et al. [421, 422],
which extends the conventional DNN by replacing one or more of its
layers with a double-projection layer and a tensor layer. In the double-
projection layer, each input vector is projected into two nonlinear sub-
spaces. In the tensor layer, two subspace projections interact with each
other and jointly predict the next layer in the overall deep architecture.
An approach is developed to map the tensor layers to the conventional
sigmoid layers so that the former can be treated and trained in a simi-
lar way to the latter. With this mapping the tensor version of the DNN
can be treated as the DNN augmented with double-projection layers
so that the backpropagation learning algorithm can be cleanly derived
and relatively easily implemented.
A related architecture to the above is the tensor version of the DSN
described in Section 6, also usefully applied to speech classification
and recognition [180, 181]. The same approach applies to mapping the
tensor layers (i.e., the upper layer in each of the many modules in the
DSN context) to the conventional sigmoid layers. Again, this mapping
simplifies the training algorithm so that it becomes not so far apart
from that for the DSN.
As discussed in Section 3.2, the concept of convolution in time
was originated in the TDNN (time-delay neural network) as a shallow
neural network [202, 382] developed during early days of speech
recognition. Only recently and when deep architectures (e.g. deep
convolutional neural network or deep CNN) were used, it has been
7.1. Acoustic modeling for speech recognition 277

found that frequency-dimension weight sharing is more effective

for high-performance phone recognition, when the HMM is used to
handle the time variability, than time-domain weight sharing as in the
previous TDNN in which the HMM was not used [1, 2, 3, 81]. These
studies also show that designing the pooling scheme in the deep CNN
to properly trade-off between invariance to vocal tract length and
discrimination among speech sounds, together with a regularization
technique of “dropout” [166], leads to even better phone recognition
performance. This set of work further points to the direction of
trading-off between trajectory discrimination and invariance expressed
in the whole dynamic pattern of speech defined in mixed time and
frequency domains using convolution and pooling. Moreover, the
most recent studies reported in [306, 307, 312] show that CNNs also
benefit large vocabulary continuous speech recognition. They further
demonstrate that multiple convolutional layers provide even more
improvement when the convolutional layers use a large number of
convolution kernels or feature maps. In particular, Sainath et al. [306]
extensively explored many variants of the deep CNN. In combination
with several novel methods the deep CNN is shown to produce state
of the art results in a few large vocabulary speech recognition tasks.
In addition to the DNN, CNN, and DSN, as well as their tensor
versions, other deep models have also been developed and reported in
the literature for speech recognition. For example, the deep-structured
CRF, which stacks many layers of CRFs, have been usefully applied
to the task of language identification [429], phone recognition [410],
sequential labeling in natural language processing [428], and confi-
dence calibration in speech recognition [423]. More recently, Demuynck
and Triefenbach [70] developed the deep GMM architecture, where the
aspects of DNNs that lead to strong performance are extracted and
applied to build hierarchical GMMs. They show that by going “deep
and wide” and feeding windowed probabilities of a lower layer of GMMs
to a higher layer of GMMs, the performance of the deep-GMM system
can be made comparable to a DNN. One advantage of staying in the
GMM space is that the decades of work in GMM adaptation and dis-
criminative learning remains applicable.
278 Selected Applications in Speech and Audio Processing

Perhaps the most notable deep architecture among all is the recur-
rent neural network (RNN) as well as its stacked or deep versions
[135, 136, 153, 279, 377]. While the RNN saw its early success in phone
recognition [304], it was not easy to duplicate due to the intricacy
in training, let alone to scale up for larger speech recognition tasks.
Learning algorithms for the RNN have been dramatically improved
since then, and much better results have been obtained recently using
the RNN [48, 134, 235], especially when the bi-directional LSTM (long
short-term memory) is used [135, 136]. The basic information flow in
the bi-directional RNN and a cell of LSTM is shown in Figures 7.3 and
7.4, respectively.
Learning the RNN parameters is known to be difficult due to van-
ishing or exploding gradients [280]. Chen and Deng [48] and Deng and

Figure 7.3: Information flow in the bi-directional RNN, with both diagrammatic
and mathematical descriptions. W’s are weight matrices, not shown but can be easily
inferred in the diagram. [after [136], @IEEE].
7.1. Acoustic modeling for speech recognition 279

Figure 7.4: Information flow in an LSTM unit of the RNN, with both diagrammatic
and mathematical descriptions. W’s are weight matrices, not shown but can easily
be inferred in the diagram. [after [136], @IEEE].

Chen [85] developed a primal-dual training method that formulates

the learning of the RNN as a formal optimization problem, where cross
entropy is maximized subject to the condition that the infinity norm of
the recurrent matrix of the RNN is less than a fixed value to guarantee
the stability of RNN dynamics. Experimental results on phone recog-
nition demonstrate: (1) the primal-dual technique is highly effective
in learning RNNs, with superior performance to the earlier heuristic
method of truncating the size of the gradient; (2) The use of a DNN
to compute high-level features of speech data to feed into the RNN
gives much higher accuracy than without using the DNN; and (3) The
280 Selected Applications in Speech and Audio Processing

accuracy drops progressively as the DNN features are extracted from

higher to lower hidden layers of the DNN.
A special case of the RNN is reservoir models or echo state networks,
where the output layers are fixed to be linear instead of nonlinear as
in the regular RNN, and where the recurrent matrices are carefully
designed but not learned. The input matrices are also fixed and not
learned, due partly to the difficulty of learning. Only the weight matri-
ces between the hidden and output layers are learned. Since the output
layer is linear, the learning is very efficient and with global optimum
achievable by a closed-form solution. But due to the fact that many
parameters are not learned, the hidden layer needs to be very large
in order to obtain good results. Triefenbach et al. [365] applied such
models to phone recognition, with reasonably good accuracy obtained.
Palangi et al. [276] presented an improved version of the reservoir
model by learning both the input and recurrent matrices which were
fixed in the previous model that makes use of the linear output (or
readout) units to simplify the learning of only the output matrix in the
RNN. Rather, a special technique is devised that takes advantage of the
linearity in the output units in the reservoir model to learn the input
and recurrent matrices. Compared with the backpropagation through
time (BPTT) algorithm commonly used in learning the general RNNs,
the proposed technique makes use of the linearity in the output units
to provide constraints among various matrices in the RNN, enabling
the computation of the gradients as the learning signal in an analytical
form instead of by recursion as in the BPTT.
In addition to the recent innovations in better architectures of deep
learning models for speech recognition reviewed above, there is also a
growing body of work on developing and implementing better nonlinear
units. Although sigmoidal and tanh functions are the most commonly
used nonlinear types in DNNs their limitations are well known. For
example, it is slow to learn the whole network due to weak gradients
when the units are close to saturation in both directions. Jaitly and
Hinton [183] appear to be the first to apply the rectified linear units
(ReLU) in the DNNs to speech recognition to overcome the weakness
of the sigmoidal units. ReLU refers to the units in a neural network
that use the activation function of f (x) = max(0, x). Dahl et al. [65]
7.1. Acoustic modeling for speech recognition 281

and Mass et al. [234] successfully applied ReLU to large vocabulary

speech recognition, with the best accuracy obtained when combining
ReLU with the “Dropout” regularization technique.
Another new type of DNN units demonstrated more recently to be
useful for speech recognition is the “maxout” units, which were used
for forming the deep maxout network as described in [244]. A deep
maxout network consists of multiple layers which generate hidden acti-
vations via the maximum or “maxout” operation over a fixed number
of weighted inputs called a “group.” This is the same operation as
the max pooling used in the CNN as discussed earlier for both speech
recognition and computer vision. The maximal value within each group
is taken as the output from the previous layer. Most recently, Zhang
et al. [441] generalize the above “maxout” units to two new types. The
“soft-maxout” type of units replace the original max operation with
the soft-max function. The second, p-norm type of units used the non-
linearity of y = xp . It is shown experimentally that the p-norm units
with p = 2 perform consistently better than the maxout, tanh, and
ReLU units. In Gulcehre et al. [138], techniques that automatically
learn the p-norm was proposed and investigated.
Finally, Srivastava et al. [350] propose yet another new type of non-
linear units, called winner-take-all units. Here, local competition among
neighboring neurons are incorporated into the otherwise regular feed-
forward architecture, which is then trained via backpropagation with
different gradients than the normal one. Winner-take-all is an inter-
esting new form of nonlinearity, and it forms groups of (typically two)
neurons where all the neurons in a group are made zero-valued except
the one with the largest value. Experiments show that the network does
not forget as much as networks with standard sigmoidal nonlinearity.
This new type of nonlinear units are yet to be evaluated in speech
recognition tasks.

7.1.7 Better optimization and regularization

Another area where significant advances are made recently in

applying deep learning to acoustic model for speech recognition is
on optimization criteria and methods, as well as on the related
282 Selected Applications in Speech and Audio Processing

regularization techniques to help prevent overfitting during the deep

network training.
One of the early studies on DNNs for speech recognition, conducted
at Microsoft Research and reported in [260], first recognizes the mis-
match between the desired error rate and the cross-entropy training
criterion in the conventional DNN training. The solution is provided by
replacing the frame-based, cross-entropy training criterion with the full-
sequence-based maximum mutual information optimization objective,
in a similar way to defining the training objective for the shallow neu-
ral network interfaced with an HMM [194]. Equivalently, this amounts
to putting the model of conditional random field (CRF) at the top of
the DNN, replacing the original softmax layer which naturally leads to
cross entropy. (Note the DNN was called the DBN in the paper). This
new sequential discriminative learning technique is developed to jointly
optimize the DNN weights, CRF transition weights, and bi-phone lan-
guage model. Importantly, the speech task is defined in TIMIT, with
the use of a simple bi-phone-gram “language” model. The simplicity of
the bi-gram language model enables the full-sequence training to carry
out without the need to use lattices, drastically reducing the training
complexity.
As another way to motivate the full-sequence training method of
[260], we note that the earlier DNN phone recognition experiments
made use of the standard frame-based objective function in static pat-
tern classification, cross-entropy, to optimize the DNN weights. The
transition parameters and language model scores were obtained from
an HMM and were trained independently of the DNN weights. However,
it has been known during the long history of the HMM research that
sequence classification criteria can be very helpful in improving speech
and phone recognition accuracy. This is because the sequence classifica-
tion criteria are more directly correlated with the performance measure
(e.g., the overall word or phone error rate) than frame-level criteria.
More specifically, the use of frame-level cross entropy to train the DNN
for phone sequence recognition does not explicitly take into account the
fact that the neighboring frames have smaller distances between the
assigned probability distributions over phone class labels. To overcome
this deficiency, one can optimize the conditional probability of the
7.1. Acoustic modeling for speech recognition 283

whole sequence of labels, given the whole visible feature utterance or

equivalently the hidden feature sequence extracted by DNN. To opti-
mize the log conditional probability on the training data, the gradi-
ent can be taken over the activation parameters, transition parame-
ters and lower-layer weights, and then pursue back-propagation of the
error defined at the sentence level. We remark that in a much earlier
study [212], combining a neural network with a CRF-like structure was
done, where the mathematical formulation appears to include CRFs as
a special case. Also, the benefit of using the full-sequence classification
criteria was shown earlier on shallow neural networks in [194, 291].
In implementing the above full-sequence learning algorithm for the
DNN system as described in [260], the DNN weights are initialized
using the frame-level cross entropy as the objective. The transition
parameters are initialized from the combination of the HMM tran-
sition matrices and the “bi-phone language” model scores, and are
then further optimized by tuning the transition features while fixing
the DNN weights before the joint optimization. Using joint optimiza-
tion with careful scheduling to reduce overfitting, it is shown that the
full-sequence training outperforms the DNN trained with frame-level
cross entropy by approximately 5% relative [260]. Without the effort
to reduce overfitting, it is found that the DNN trained with MMI is
much more prone to overfitting than that trained with frame-level cross
entropy. This is because the correlations across frames in speech tend
to be different among the training, development, and test data. Impor-
tantly, such differences do not show when frame-based objective func-
tions are used for training.
For large vocabulary speech recognition where more complex lan-
guage models are in use, the optimization methods for full-sequence
training of the DNN–HMM are much more sophisticated. Kingsbury
et al. [195] reported the first success of such training using parallel,
second-order, Hessian-free optimization techniques, which are carefully
implemented for large vocabulary speech recognition. Sainath et al.
[305] improved and speeded up the Hessian-free techniques by reduc-
ing the number of Krylov subspace solver iterations [378], which are
used for implicit estimation of the Hessian. They also use sampling
284 Selected Applications in Speech and Audio Processing

methods to decrease the amount of training data to speed up the train-

ing. While the batch-mode, second-order Hessian-free techniques prove
successful for full-sequence training of large-scale DNN–HMM systems,
the success of the first-order stochastic gradient descent methods is also
reported recently [353]. It is found that heuristics are needed to handle
the problem of lattice sparseness. That is, the DNN must be adjusted to
the updated numerator lattices by additional iterations of frame-based
cross-entropy training. Further, artificial silence arcs need to be added
to the denominator lattices, or the maximum mutual information objec-
tive function needs to be smoothed with the frame-based cross entropy
objective. The conclusion is that for large vocabulary speech recog-
nition tasks with sparse lattices, the implementation of the sequence
training requires much greater engineering skills than the small tasks
such as reported in [260], although the objective function as well as the
gradient derivation are essentially the same. Similar conclusions are
reached by Vesely et al. [374] when carrying out full-sequence training
of DNN–HMMs for large-vocabulary speech recognition. However, dif-
ferent heuristics from [353] are shown to be effective in the training.
Separately, Wiesler et al. [390] investigated the Hessian-free optimiza-
tion method for training the DNN with the cross-entropy objective and
empirically analyzed the properties of the method. And finally, Dognin
and Goel [113] combined stochastic average gradient and Hessian-free
optimization for sequence training of deep neural networks with suc-
cess in that the training procedure converges in about half the time
compared with the full Hessian-free sequence training.
For large DNN–HMM systems with either frame-level or sequence-
level optimization objectives, speeding up the training is essential
to take advantage of large amounts of training data and of large
model sizes. In addition to the methods described above, Dean et al.
[69] reported the use of the asynchronous stochastic gradient descent
(ASGD) method, the adaptive gradient descent (Adagrad) method, and
the large-scale limited-memory BFGS (L-BFGS) method for very large
vocabulary speech recognition. Sainath et al. [312] provided a review
of a wide range of optimization methods for speeding up the training
of DNN-based systems for large speech recognition tasks.
7.1. Acoustic modeling for speech recognition 285

In addition to the advances described above focusing on optimiza-

tion with the fully supervised learning paradigm, where all train-
ing data contain the label information, the semi-supervised training
paradigm is also exploited for learning DNN–HMM systems for speech
recognition. Liao et al. [223] reported the exploration of using semi-
supervised training on the DNN–HMM system for the very challenging
task of recognizing YouTube speech. The main technique is based on the
use of “island of confidence” filtering heuristics to select useful training
segments. Separately, semi-supervised training of DNNs is explored by
Vesely et al. [374], where self-training strategies are used as the basis
for data selection using both the utterance-level and frame-level con-
fidences. Frame-selection based on per-frame confidences derived from
confusion in a lattice is found beneficial. Huang et al. [176] reported
another variant of semi-supervised training technique in which multi-
system combination and confidence recalibration is applied to select
the training data. Further, Thomas et al. [362] overcome the problem
of lacking sufficient training data for acoustic modeling in a number of
low-resource scenarios. They make use of transcribed multilingual data
and semi-supervised training to build the proposed feature front-ends
for subsequent speech recognition.
Finally, we see important progress in deep learning based speech
recognition in recent years with the introduction of new regulariza-
tion methods based on “dropout” originally proposed by Hinton et al.
[166]. Overfitting is very common in DNN training and co-adaptation is
prevalent within the DNN with multiple activations adapting together
to explain input acoustic data. Dropout is a technique to limit co-
adaptation. It operates as follows. On each training instance, each hid-
den unit is randomly omitted with a fixed probability (e.g., p = 0.5).
Then, decoding is done normally except with straightforward scaling
of the DNN weights (by a factor of 1 − p). Alternatively, the scaling of
the DNN weights can be done during training [by a factor of 1/(1 − p)]
rather than in decoding. The benefits of dropout regularization for
training DNNs are to make a hidden unit in the DNN act strongly by
itself without relying on others, and to serve a way to do model averag-
ing of different networks. These benefits are most pronounced when the
training data is limited, or when the DNN size is disproportionally large
286 Selected Applications in Speech and Audio Processing

with respect to the size of the training data. Dahl et al. [65] applied
dropout in conjunction with the ReLU units and to only the top few
layers of a fully-connected DNN. Seltzer and Yu [325] applied it to noise
robust speech recognition. Deng et al. [81], on the other hand, applied
dropout to all layers of a deep convolutional neural network, including
both the top fully connected DNN layers and the bottom locally con-
nected CNN layer and the pooling layer. It is found that the dropout
rate need to be substantially smaller for the convolutional layer.
Subsequent work on applying dropout includes the study by Miao
and Metze [243], where DNN-based speech recognition is constrained
by low resources with sparse training data. Most recently, Sainath et al.
[306] combined dropout with a number of novel techniques described
in this section (including the use of deep CNNs, Hessian-free sequence
learning, the use of ReLU units, and the use of joint fMLLR and filter-
bank features, etc.) to obtain state of the art results on several large
vocabulary speech recognition tasks.
As a summary, the initial success of deep learning methods for
speech analysis and recognition reported around 2010 has come a long
way over the past three years. An explosive growth in the work and
publications on this topic has been observed, and huge excitement has
been ignited within the speech recognition community. We expect that
the growth in the research on deep learning based speech recognition
will continue, at least in the near future. It is also fair to say that the
continuing large-scale success of deep learning in speech recognition as
surveyed in this chapter (up to the ASRU-2013 time frame) is a key
stimulant to the large-scale exploration and applications of the deep
learning methods to other areas, which we will survey in Sections 8–11.

7.2 Speech synthesis

In addition to speech recognition, the impact of deep learning has

recently spread to speech synthesis, aimed to overcome the limitations
of the conventional approach in statistical parametric synthesis based
on Gaussian-HMM and decision-tree-based model clustering. The goal
of speech synthesis is to generate speech sounds directly from text and
7.2. Speech synthesis 287

possibly with additional information. The first set of papers appeared at

ICASSP, May 2013, where four different deep learning approaches are
reported to improve the traditional HMM-based statistical paramet-
ric speech synthesis systems built based on “shallow” speech models,
which we briefly review here after providing appropriate background
information.
Statistical parametric speech synthesis emerged in the mid-1990s,
and is currently the dominant technology in speech synthesis. See a
recent overview in [364]. In this approach, the relationship between
texts and their acoustic realizations are modeled using a set of
stochastic generative acoustic models. Decision tree-clustered context-
dependent HMMs with a Gaussian distribution as the output of an
HMM state are the most popular generative acoustic model used. In
such HMM-based speech synthesis systems, acoustic features including
the spectra, excitation and segment durations of speech are modeled
simultaneously within a unified context-dependent HMM framework.
At the synthesis time, a text analysis module extracts a sequence of
contextual factors including phonetic, prosodic, linguistic, and gram-
matical descriptions from an input text to be synthesized. Given
the sequence of contextual factors, a sentence-level context-dependent
HMM corresponding to the input text is composed, where its model
parameters are determined by traversing the decision trees. The acous-
tic features are predicted so as to maximize their output probabili-
ties from the sentence HMM under the constraints between static and
dynamic features. Finally, the predicted acoustic features are sent to a
waveform synthesis module to reconstruct the speech waveforms. It
has been known for many years that the speech sounds generated
by this standard approach are often muffled compared with natural
speech. The inadequacy of acoustic modeling based on the shallow-
structured HMM is conjectured to be one of the reasons. Several very
recent studies have adopted deep learning approaches to overcome such
deficiency. One significant advantage of deep learning techniques is
their strong ability to represent the intrinsic correlation or mapping
relationship among the units of a high-dimensional stochastic vector
using a generative (e.g., the RBM and DBN discussed in Section 3.2)
or discriminative (e.g., the DNN discussed in Section 3.3) modeling
288 Selected Applications in Speech and Audio Processing

framework. The deep learning techniques are thus expected to help the
acoustic modeling aspect of speech synthesis in overcoming the limita-
tions of the conventional shallow modeling approach.
A series of studies are carried out recently on ways of overcoming
the above limitations using deep learning methods, inspired partly by
the intrinsically hierarchical processes in human speech production
and the successful applications of a number of deep learning methods
in speech recognition as reviewed earlier in this chapter. In Ling
et al. [227, 229], the RBM and DBN as generative models are used
to replace the traditional Gaussian models, achieving significant
quality improvement, in both subjective and objective measures,
of the synthesized voice. In the approach developed in [190], the
DBN as a generative model is used to represent joint distribution of
linguistic and acoustic features. Both the decision trees and Gaussian
models are replaced by the DBN. The method is very similar to that
used for generating digit images by the DBN, where the issue of
temporal sequence modeling specific to speech (non-issue for image)
is by-passed via the use of the relatively large, syllable-sized units in
speech synthesis. On the other hand, in contrast to the generative
deep models (RBMs and DBNs) exploited above, the study reported
in [435] makes use of the discriminative model of the DNN to represent
the conditional distribution of the acoustic features given the linguistic
features. Finally, in [115], the discriminative model of the DNN is used
as a feature extractor that summarizes high-level structure from the
raw acoustic features. Such DNN features are then used as the input
for the second stage for the prediction of prosodic contour targets
from contextual features in the full speech synthesis system.
The application of deep learning to speech synthesis is in its infancy,
and much more work is expected from that community in the near
future.

7.3 Audio and music processing

Similar to speech recognition but to a less extent, in the area of audio

and music processing, deep learning has also become of intense interest
7.3. Audio and music processing 289

but only quite recently. As an example, the first major event of deep
learning for speech recognition took place in 2009, followed by a series of
events including a comprehensive tutorial on the topic at ICASSP-2012
and with the special issue at IEEE Transactions on Audio, Speech, and
Language Processing, the premier publication for speech recognition,
in the same year. The first major event of deep learning for audio and
music processing appears to be the special session at ICASSP-2014,
titled Deep Learning for Music [14].
In the general field of audio and music processing, the impacted
areas by deep learning include mainly music signal processing and music
information retrieval [15, 22, 141, 177, 178, 179, 319]. Deep learning
presents a unique set of challenges in these areas. Music audio signals
are time series where events are organized in musical time, rather than
in real time, which changes as a function of rhythm and expression. The
measured signals typically combine multiple voices that are synchro-
nized in time and overlapping in frequency, mixing both short-term and
long-term temporal dependencies. The influencing factors include musi-
cal tradition, style, composer and interpretation. The high complexity
and variety give rise to the signal representation problems well-suited
to the high levels of abstraction afforded by the perceptually and bio-
logically motivated processing techniques of deep learning.
In the early work on audio signals as reported by Lee et al. [215]
and their follow-up work, the convolutional structure is imposed on
the RBM while building up a DBN. Convolution is made in time by
sharing weights between hidden units in an attempt to detect the same
“invariant” feature over different times. Then a max-pooling operation
is performed where the maximal activations over small temporal neigh-
borhoods of hidden units are obtained, inducing some local temporal
invariance. The resulting convolutional DBN is applied to audio as well
as speech data for a number of tasks including music artist and genre
classification, speaker identification, speaker gender classification, and
phone classification, with promising results presented.
The RNN has also been recently applied to music processing appli-
cations [22, 40, 41], where the use of ReLU hidden units instead of
logistic or tanh nonlinearities are explored in the RNN. As reviewed in
290 Selected Applications in Speech and Audio Processing

Section 7.2, ReLU units compute y = max(x, 0), and lead to sparser
gradients, less diffusion of credit and blame in the RNN, and faster
training. The RNN is applied to the task of automatic recognition of
chords from audio music, an active area of research in music information
retrieval. The motivation of using the RNN architecture is its power
in modeling dynamical systems. The RNN incorporates an internal
memory, or hidden state, represented by a self-connected hidden layer
of neurons. This property makes them well suited to model temporal
sequences, such as frames in a magnitude spectrogram or chord labels
in a harmonic progression. When well trained, the RNN is endowed
with the power to predict the output at the next time step given the
previous ones. Experimental results show that the RNN-based auto-
matic chord recognition system is competitive with existing state-of-
the-art approaches [275]. The RNN is capable of learning basic musical
properties such as temporal continuity, harmony and temporal dynam-
ics. It can also efficiently search for the most musically plausible chord
sequences when the audio signal is ambiguous, noisy or weakly discrim-
inative.
A recent review article by Humphrey et al. [179] provides a detailed
analysis on content-based music informatics, and in particular on why
the progress is decelerating throughout the field. The analysis con-
cludes that hand-crafted feature design is sub-optimal and unsustain-
able, that the power of shallow architectures is fundamentally limited,
and that short-time analysis cannot encode musically meaningful struc-
ture. These conclusions motivate the use of deep learning methods
aimed at automatic feature learning. By embracing feature learning, it
becomes possible to optimize a music retrieval system’s internal feature
representation or discovering it directly, since deep architectures are
especially well-suited to characterize the hierarchical nature of music.
Finally, we review the very recent work by van den Oord, et al. [371]
on content-based music recommendation using deep learning methods.
Automatic music recommendation has become an increasingly signifi-
cant and useful technique in practice. Most recommender systems rely
on collaborative filtering, suffering from the cold start problem where
it fails when no usage data is available. Thus, collaborative filtering is
7.3. Audio and music processing 291

not effective for recommending new and unpopular songs. Deep learning
methods power the latent factor model for recommendation, which pre-
dicts the latent factors from music audio when they cannot be obtained
from usage data. A traditional approach using a bag-of-words represen-
tation of the audio signals is compared with deep CNNs with rigorous
evaluation made. The results show highly sensible recommendations
produced by the predicted latent factors using deep CNNs. The study
demonstrates that a combination of convolutional neural networks and
richer audio features lead to such promising results for content-based
music recommendation.
Like speech recognition and speech synthesis, much more work is
expected from the music and audio signal processing community in the
near future.
 8
Selected Applications in Language
Modeling and Natural Language Processing

Research in language, document, and text processing has seen

increasing popularity recently in the signal processing community,
and has been designated as one of the main focus areas by the IEEE
Signal Processing Society’s Speech and Language Processing Technical
Committee. Applications of deep learning to this area started with
language modeling (LM), where the goal is to provide a probability
to any arbitrary sequence of words or other linguistic symbols (e.g.,
letters, characters, phones, etc.). Natural language processing (NLP)
or computational linguistics also deals with sequences of words or
other linguistic symbols, but the tasks are much more diverse (e.g.,
translation, parsing, text classification, etc.), not focusing on providing
probabilities for linguistic symbols. The connection is that LM is
often an important and very useful component of NLP systems.
Applications to NLP is currently one of the most active areas in
deep learning research, and deep learning is also considered as one
promising direction by the NLP research community. However, the
intersection between the deep learning and NLP researchers is so far
not nearly as large as that for the application areas of speech or vision.
This is partly because the hard evidence for the superiority of deep

292
8.1. Language modeling 293

learning over the current state of the art NLP methods has not been
as strong as speech or visual object recognition.

8.1 Language modeling

Language models (LMs) are crucial part of many successful applica-

tions, such as speech recognition, text information retrieval, statistical
machine translation and other tasks of NLP. Traditional techniques for
estimating the parameters in LMs are based on N-gram counts. Despite
known weaknesses of N -grams and huge efforts of research communities
across many fields, N -grams remained the state-of-the-art until neural
network and deep learning based methods were shown to significantly
lower the perplexity of LMs, one common (but not ultimate) measure of
the LM quality, over several standard benchmark tasks [245, 247, 248].
Before we discuss neural network based LMs, we note the use of
hierarchical Bayesian priors in building up deep and recursive struc-
ture for LMs [174]. Specifically, Pitman-Yor process is exploited as the
Bayesian prior, from which a deep (four layers) probabilistic genera-
tive model is built. It offers a principled approach to LM smoothing
by incorporating the power-law distribution for natural language. As
discussed in Section 3, this type of prior knowledge embedding is more
readily achievable in the generative probabilistic modeling setup than
in the discriminative neural network based setup. The reported results
on LM perplexity reduction are not nearly as strong as that achieved
by the neural network based LMs, which we discuss next.
There has been a long history [19, 26, 27, 433] of using (shallow)
feed-forward neural networks in LMs, called the NNLM. The use of
DNNs in the same way for LMs appeared more recently in [8]. An LM
is a function that captures the salient statistical characteristics of the
distribution of sequences of words in natural language. It allows one to
make probabilistic predictions of the next word given preceding ones.
An NNLM is one that exploits the neural network’s ability to learn
distributed representations in order to reduce the impact of the curse
of dimensionality. The original NNLM, with a feed-forward neural net-
work structure works as follows: the input of the N-gram NNLM is
294 Language Modeling and Natural Language Processing

formed by using a fixed length history of N − 1 words. Each of the

previous N − 1 words is encoded using the very sparse 1-of-V coding,
where V is the size of the vocabulary. Then, this 1-of-V orthogonal rep-
resentation of words is projected linearly to a lower dimensional space,
using the projection matrix shared among words at different positions
in the history. This type of continuous-space, distributed representation
of words is called “word embedding,” very different from the common
symbolic or localist presentation [26, 27]. After the projection layer,
a hidden layer with nonlinear activation function, which is either a
hyperbolic tangent or a logistic sigmoid, is used. An output layer of
the neural network then follows the hidden layer, with the number of
output units equal to the size of the full vocabulary. After the network
is trained, the output layer activations represent the “N -gram” LM’s
probability distribution.
The main advantage of NNLMs over the traditional counting-based
N -gram LMs is that history is no longer seen as exact sequence of N −1
words, but rather as a projection of the entire history into some lower
dimensional space. This leads to a reduction of the total number of
parameters in the model that have to be trained, resulting in automatic
clustering of similar histories. Compared with the class-based N -gram
LMs, the NNLMs are different in that they project all words into the
same low dimensional space, in which there can be many degrees of
similarity between words. On the other hand, NNLMs have much larger
computational complexity than N -gram LMs.
Let’s look at the strengths of the NNLMs again from the view-
point of distributed representations. A distributed representation of a
symbol is a vector of features which characterize the meaning of the
symbol. Each element in the vector participates in representing the
meaning. With an NNLM, one relies on the learning algorithm to dis-
cover meaningful, continuous-valued features. The basic idea is to learn
to associate each word in the dictionary with a continuous-valued vec-
tor representation, which in the literature is called a word embedding,
where each word corresponds to a point in a feature space. One can
imagine that each dimension of that space corresponds to a semantic
or grammatical characteristic of words. The hope is that functionally
8.1. Language modeling 295

similar words get to be closer to each other in that space, at least along
some directions. A sequence of words can thus be transformed into a
sequence of these learned feature vectors. The neural network learns to
map that sequence of feature vectors to the probability distribution over
the next word in the sequence. The distributed representation approach
to LMs has the advantage that it allows the model to generalize well to
sequences that are not in the set of training word sequences, but that
are similar in terms of their features, i.e., their distributed represen-
tation. Because neural networks tend to map nearby inputs to nearby
outputs, the predictions corresponding to word sequences with similar
features are mapped to similar predictions.
The above ideas of NNLMs have been implemented in various
studies, some involving deep architectures. The idea of structuring
hierarchically the output of an NNLM in order to handle large
vocabularies was introduced in [18, 262]. In [252], the temporally
factored RBM was used for language modeling. Unlike the traditional
N -gram model, the factored RBM uses distributed representations
not only for context words but also for the words being predicted.
This approach is generalized to deeper structures as reported in [253].
Subsequent work on NNLM with “deep” architectures can be found in
[205, 207, 208, 245, 247, 248]. As an example, Le et al. [207] describes
an NNLM with structured output layer (SOUL–NNLM) where the pro-
cessing depth in the LM is focused in the neural network’s output rep-
resentation. Figure 8.1 illustrates the SOUL-NNLM architecture with
hierarchical structure in the output layers of the neural network, which
shares the same architecture with the conventional NNLM up to the
hidden layer. The hierarchical structure for the network’s output vocab-
ulary is in the form of a clustering tree, shown to the right of Figure 8.1,
where each word belongs to only one class and ends in a single leaf node
of the tree. As a result of the hierarchical structure, the SOUL–NNLM
enables the training of the NNLM with a full, very large vocabulary.
This gives advantages over the traditional NNLM which requires short-
lists of words in order to carry out the efficient computation in training.
As another example neural-network-based LMs, the work described
in [247, 248] and [245] makes use of RNNs to build large scale language
296 Language Modeling and Natural Language Processing

Figure 8.1: The SOUL–NNLM architecture with hierarchical structure in the out-
put layers of the neural network [after [207], @IEEE].

models, called RNNLMs. The main difference between the feed-forward

and the recurrent architecture for LMs is different ways of representing
the word history. For feed-forward NNLM, the history is still just pre-
vious several words. But for the RNNLM, an effective representation
of history is learned from the data during training. The hidden layer of
RNN represents all previous history and not just N − 1 previous words,
thus the model can theoretically represent long context patterns. A fur-
ther important advantage of the RNNLM over the feed-forward coun-
terpart is the possibility to represent more advanced patterns in the
word sequence. For example, patterns that rely on words that could
have occurred at variable positions in the history can be encoded much
more efficiently with the recurrent architecture. That is, the RNNLM
can simply remember some specific word in the state of the hidden
layer, while the feed-forward NNLM would need to use parameters for
each specific position of the word in the history.
The RNNLM is trained using the algorithm of back-propagation
through time; see details in [245], which provided Figure 8.2 to show
during training how the RNN unfolds as a deep feed-forward network
(with three time steps back in time).
8.1. Language modeling 297

Figure 8.2: During the training of RNNLMs, the RNN unfolds into a deep feed-
forward network; based on Figure 3.2 of [245].

The training of the RNNLM achieves stability and fast convergence,

helped by capping the growing gradient in training RNNs. Adaptation
schemes for the RNNLM are also developed by sorting the training
data with respect to their relevance and by training the model during
processing of the test data. Empirical comparisons with other state-of-
the-art counting-based N -gram LMs show much better performance of
RNNLM in the perplexity measure, as reported in [247, 248] and [245].
298 Language Modeling and Natural Language Processing

A separate work on applying RNN to an LM with the unit

of characters instead of words can be found in [153, 357]. Many
interesting properties such as predicting long-term dependencies (e.g.,
making open and closing quotes in a paragraph) are demonstrated.
However, the usefulness of characters instead of words as units in
practical applications is not clear because the word is such a powerful
representation for natural language. Changing words to characters in
LMs may limit most practical application scenarios and the training
become more difficult. Word-level models currently remain superior.
In the most recent work, Mnih and Teh [255] and Mnih and
Kavukcuoglu [254] have developed a fast and simple training algorithm
for NNLMs. Despite their superior performance, NNLMs have been
used less widely than standard N -gram LMs due to the much longer
training time. The reported algorithm makes use of a method called
noise-contrastive estimation or NCE [139] to achieve much faster train-
ing for NNLMs, with time complexity independent of the vocabulary
size; hence a flat instead of tree-structured output layer in the NNLM
is used. The idea behind NCE is to perform nonlinear logistic regres-
sion to discriminate between the observed data and some artificially
generated noise. That is, to estimate parameters in a density model of
observed data, we can learn to discriminate between samples from the
data distribution and samples from a known noise distribution. As an
important special case, NCE is particularly attractive for unnormalized
distributions (i.e., free from partition functions in the denominator). In
order to apply NCE to train NNLMs efficiently, Mnih and Teh [255]
and Mnih and Kavukcuoglu [254] first formulate the learning problem
as one which takes the objective function as the distribution of the word
in terms of a scoring function. The NNLM then can be viewed as a way
to quantify the compatibility between the word history and a candidate
next word using the scoring function. The objective function for train-
ing the NNLM thus becomes exponentiation of the scoring function,
normalized by the same constant over all possible words. Removing
the costly normalization factor, NCE is shown to speed up the NNLM
training over an order of magnitude.
A similar concept to NCE is used in the recent work of [250], which
is called negative sampling. This is applied to a simplified version of
8.2. Natural language processing 299

an NNLM, for the purpose of constructing word embedding instead

of computing probabilities of word sequences. Word embedding is an
important concept for NLP applications, which we discuss next.

8.2 Natural language processing

Machine learning has been a dominant tool in NLP for many years.
However, the use of machine learning in NLP has been mostly limited to
numerical optimization of weights for human designed representations
and features from the text data. The goal of deep or representation
learning is to automatically develop features or representations from
the raw text material appropriate for a wide range of NLP tasks.
Recently, neural network based deep learning methods have
been shown to perform well on various NLP tasks such as language
modeling, machine translation, part-of-speech tagging, named entity
recognition, sentiment analysis, and paraphrase detection. The most
attractive aspect of deep learning methods is their ability to perform
these tasks without external hand-designed resources or time-intensive
feature engineering. To this end, deep learning develops and makes use
an important concept called “embedding,” which refers to the represen-
tation of symbolic information in natural language text at word-level,
phrase-level, and even sentence-level in terms of continuous-valued
vectors.
The early work highlighting the importance of word embedding
came from [62], [367], and [63], although the original form came from
[26] as a side product of language modeling. Raw symbolic word rep-
resentations are transformed from the sparse vectors via 1-of-V coding
with a very high dimension (i.e., the vocabulary size V or its square or
even its cubic) into low-dimensional, real-valued vectors via a neural
network and then used for processing by subsequent neural network lay-
ers. The key advantage of using the continuous space to represent words
(or phrases) is its distributed nature, which enables sharing or grouping
the representations of words with a similar meaning. Such sharing is
not possible in the original symbolic space, constructed by 1-of-V cod-
ing with a very high dimension, for representing words. Unsupervised
300 Language Modeling and Natural Language Processing

learning is used where “context” of the word is used as the learning

signal in neural networks. Excellent tutorials were recently given by
Socher et al. [338, 340] to explain how the neural network is trained
to perform word embedding. More recent work proposes new ways of
learning word embeddings that better capture the semantics of words
by incorporating both local and global document contexts and better
account for homonymy and polysemy by learning multiple embeddings
per word [169]. Also, there is strong evidence that the use of RNNs can
also provide empirically good performance in learning word embeddings
[245]. While the use of NNLMs, whose aim is to predict the future words
in context, also induces word embeddings as its by-product, much sim-
pler ways of achieving the embeddings are possible without the need to
do word prediction. As shown by Collobert and Weston [62], the neural
networks used for creating word embeddings need much smaller output
units than the huge size typically required for NNLMs.
In the same early paper on word embedding, Collobert and Weston
[62] developed and employed a convolutional network as the common
model to simultaneously solve a number of classic problems includ-
ing part-of-speech tagging, chunking, named entity tagging, semantic
role identification, and similar word identification. More recent work
reported in [61] further developed a fast, purely discriminative approach
for parsing based on the deep recurrent convolutional architecture. Col-
lobert et al. [63] provide a comprehensive review on ways of applying
unified neural network architectures and related deep learning algo-
rithms to solve NLP problems from “scratch,” meaning that no tradi-
tional NLP methods are used to extract features. The theme of this
line of work is to avoid task-specific, “man-made” feature engineering
while providing versatility and unified features constructed automati-
cally from deep learning applicable to all natural language processing
tasks. The systems described in [63] automatically learn internal repre-
sentations or word embedding from vast amounts of mostly unlabeled
training data while performing a wide range of NLP tasks.
The recent work by Mikolov et al. [246] derives word embeddings
by simplifying the NNLM described in Section 8.1. It is found that
the NNLM can be successfully trained in two steps. First, continuous
word vectors are learned using a simple model which eliminates the
8.2. Natural language processing 301

Figure 8.3: The CBOW architecture (a) on the left, and the Skip-gram architecture
(b) on the right. [after [246], @ICLR].

nonlinearity in the upper neural network layer and share the projec-
tion layer for all words. And second, the N -gram NNLM is trained
on top of the word vectors. So, after removing the second step in the
NNLM, the simple model is used to learn word embeddings, where the
simplicity allows the use of very large amount of data. This gives rise
to a word embedding model called Continuous Bag-of-Words Model
(CBOW), as shown in Figure 8.3a. Further, since the goal is no longer
computing probabilities of word sequences as in LMs, the word embed-
ding system here is made more effective by not only to predict the
current word based on the context but also to perform inverse pre-
diction known as “Skip-gram” model, as shown in Figure 8.3b. In
the follow-up work [250] by the same authors, this word embedding
system including the Skip-gram model is extended by a much faster
learning method called negative sampling, similar to NCE discussed in
Section 8.1.
In parallel with the above development, Mnih and Kavukcuoglu
[254] demonstrate that NCE training of lightweight word embedding
302 Language Modeling and Natural Language Processing

models is a highly efficient way of learning high-quality word represen-

tations, much like the somewhat earlier lightweight LMs developed by
Mnih and Teh [255] described in Section 8.1. Consequently, results that
used to require very considerable hardware and software infrastructure
can now be obtained on a single desktop with minimal programming
effort and using less time and data. This most recent work also shows
that for representation learning, only five noise samples in NCE can be
sufficient for obtaining strong results for word embedding, much fewer
than that required for LMs. The authors also used an “inversed lan-
guage model” for computing word embeddings, similar to the way in
which the Skip-gram model is used in [250].
Huang et al. [169] recognized the limitation of the earlier work on
word embeddings in that these models were built with only local con-
text and one representation per word. They extended the local context
models to one that can incorporate global context from full sentences
or the entire document. This extended models accounts for homonymy
and polysemy by learning multiple embeddings for each word. An illus-
tration of this model is shown in Figure 8.4. In the earlier work by the
same research group [344], a recursive neural network with local con-
text was developed to build a deep architecture. The network, despite
missing global context, was already shown to be capable of successful

Figure 8.4: The extended word-embedding model using a recursive neural network
that takes into account not only local context but also global context. The global
context is extracted from the document and put in the form of a global semantic
vector, as part of the input into the original word-embedding model with local
context. Taken from Figure 1 of [169]. [after [169], @ACL].
8.2. Natural language processing 303

merging of natural language words based on the learned semantic trans-

formations of their original features. This deep learning approach pro-
vided an excellent performance on natural language parsing. The same
approach was also demonstrated to be reasonably successful in pars-
ing natural scene images. In related studies, a similar recursive deep
architecture is used for paraphrase detection [346], and for predicting
sentiment distributions from text [345].
We now turn to selected applications of deep learning methods
including the use of neural network architectures and word embed-
dings to practically useful NLP tasks. Machine translation is one of
such tasks, pursued by NLP researchers for many years based typically
on shallow statistical models. The work described in [320] are perhaps
the first comprehensive report on the successful application of neural-
network-based language models with word embeddings, trained on a
GPU, for large machine translation tasks. They address the problem of
high computation complexity, and provide a solution that allows train-
ing 500 million words with 20 hours. Strong results are reported, with
perplexity down from 71 to 60 in LMs and the corresponding BLEU
score gained by 1.8 points using the neural-network-based language
models with word embeddings compared with the best back-off LM.
A more recent study on applying deep learning methods to machine
translation appears in [121, 123], where the phrase-translation compo-
nent, rather than the LM component in the machine translation system
is replaced by the neural network models with semantic word embed-
dings. As shown in Figure 8.5 for the architecture of this approach,
a pair of source (denoted by f ) and target (denoted by e) phrases
are projected into continuous-valued vector representations in a low-
dimensional latent semantic space (denoted by the two y vectors).Then
their translation score is computed by the distance between the pair in
this new space. The projection is performed by two deep neural net-
works (not shown here) whose weights are learned on parallel training
data. The learning is aimed to directly optimize the quality of end-
to-end machine translation results. Experimental evaluation has been
performed on two standard Europarl translation tasks used by the NLP
community, English–French and German–English. The results show
that the new semantic-based phrase translation model significantly
304 Language Modeling and Natural Language Processing

Figure 8.5: Illustration of the basic approach reported in [122] for machine trans-
lation. Parallel pairs of source (denoted by f) and target (denoted by e) phrases
are projected into continuous-valued vector representations (denoted by the two y
vectors), and their translation score is computed by the distance between the pair in
this continuous space. The projection is performed by deep neural networks (denoted
by the two arrows) whose weights are learned on parallel training data. [after [121],
@NIPS].

improves the performance of a state-of-the-art phrase-based statisti-

cal machine translation system, leading to a gain close to 1.0 BLEU
point.
A related approach to machine translation was developed by
Schwenk [320]. The estimation of the translation model probabilities of
a phrase-based machine translation system is carried out using neural
networks. The translation probability of phrase pairs is learned using
continuous-space representations induced by neural networks. A sim-
plification is made that decomposes the translation probability of a
phrase or a sentence to a product of n-gram probabilities as in a stan-
dard n-gram language model. No joint representations of a phrase in
the source language and the translated version in the target language
are exploited as in the approach reported by Gao et al. [122, 123].
Yet another deep learning approach to machine translation
appeared in [249]. As in other approaches, a corpus of words in one
language are compared with the same corpus of words translated into
another, and words and phrases in such bilingual data that share similar
statistical properties are considered equivalent. A new technique is
8.2. Natural language processing 305

proposed that automatically generates dictionaries and phrase tables

that convert one language into another. It does not rely on versions of
the same document in different languages. Instead, it uses data mining
techniques to model the structure of a source language and then com-
pares it to the structure of the target language. The technique is shown
to translate missing word and phrase entries by learning language struc-
tures based on large monolingual data and mapping between languages
from small bilingual data. It is based on vector-valued word embed-
dings as discussed earlier in this chapter and it learns a linear mapping
between vector spaces of source and target languages.
An earlier study on applying deep learning techniques with DBNs
was provided in [111] to attack a machine transliteration problem,
a much easier task than machine translation. This type of deep
architectures and learning may be generalized to the more difficult
machine translation problem but no follow-up work has been reported.
As another early NLP application, Sarikaya et al. [318] applied DNNs
(called DBNs in the paper) to perform a natural language call–routing
task. The DNNs use unsupervised learning to discover multiple layers
of features that are then used to optimize discrimination. Unsupervised
feature discovery is found to make DBNs far less prone to overfitting
than the neural networks initialized with random weights. Unsuper-
vised learning also makes it easier to train neural networks with many
hidden layers. DBNs are found to produce better classification results
than several other widely used learning techniques, e.g., maximum
entropy and boosting based classifiers.
One most interesting NLP task recently tackled by deep learn-
ing methods is that of knowledge base (ontology) completion, which
is instrumental in question-answering and many other NLP applica-
tions. An early work in this space came from [37], where a process is
introduced to automatically learn structured distributed embeddings
of knowledge bases. The proposed representations in the continuous-
valued vector space are compact and can be efficiently learned from
large-scale data of entities and relations. A specialized neural network
architecture, a generalization of “Siamese” network, is used. In the
follow-up work that focuses on multi-relational data [36], the semantic
matching energy model is proposed to learn vector representations for
306 Language Modeling and Natural Language Processing

both entities and relations. More recent work [340] adopts an alterna-
tive approach, based on the use of neural tensor networks, to attack
the problem of reasoning over a large joint knowledge graph for rela-
tion classification. The knowledge graph is represented as triples of a
relation between two entities, and the authors aim to develop a neu-
ral network model suitable for inference over such relationships. The
model they presented is a neural tensor network, with one layer only.
The network is used to represent entities in a fixed-dimensional vectors,
which are created separately by averaging pre-trained word embedding
vectors. It then learn the tensor with the newly added relationship ele-
ment that describes the interactions among all the latent components
in each of the relationships. The neural tensor network can be visu-
alized in Figure 8.6, where each dashed box denotes one of the two
slices of the tensor. Experimentally, the paper [340] shows that this
tensor model can effectively classify unseen relationships in WordNet
and FreeBase.
As the final example of deep learning applied successfully to NLP,
we discuss here sentiment analysis applications based on recursive deep

Figure 8.6: Illustration of the neural tensor network described in [340], with two
relationships shown as two slices in the tensor. The tensor is denoted by W [1:2] . The
network contains a bilinear tensor layer that directly relates the two entity vectors
(shown as e1 and e 2 ) across three dimensions. Each dashed box denotes one of the
two slices of the tensor. [after [340], @NIPS].
8.2. Natural language processing 307

models published recently by Socher et al. [347]. Sentiment analysis is

a task that is aimed to estimate the positive or negative opinion by an
algorithm based on input text information. As we discussed earlier in
this chapter, word embeddings in the semantic space achieved by neural
network models have been very useful but it is difficult for them to
express the meaning of longer phrases in a principled way. For sentiment
analysis with the input data from typically many words and phrases,
the embedding model requires the compositionality properties. To this
end, Socher et al. [347] developed the recursive neural tensor network,
where each layer is constructed similarly to that of the neural tensor
network described in [340] with an illustration shown in Figure 8.6.
The recursive construction of the full network exhibiting properties
of compositionality follows that of [344] for the regular, non-tensor
network. When trained on a carefully constructed sentiment analysis
database, the recursive neural tensor network is shown to outperform all
previous methods on several metrics. The new model pushes the state of
the art in single sentence positive/negative classification accuracy from
80% up to 85.4%. The accuracy of predicting fine-grained sentiment
labels for all phrases reaches 80.7%, an improvement of 9.7% over bag-
of-features baselines.
 9
Selected Applications in Information Retrieval

9.1 A brief introduction to information retrieval

Information retrieval (IR) is a process whereby a user enters a query

into the automated computer system that contains a collection of many
documents with the goal of obtaining a set of most relevant documents.
Queries are formal statements of information needs, such as search
strings in web search engines. In IR, a query does not uniquely identify
a single document in the collection. Instead, several documents may
match the query with different degrees of relevancy.
A document, sometimes called an object as a more general term
which may include not only a text document but also an image, audio
(music or speech), or video, is an entity that contains information
and represented as an entry in a database. In this section, we limit
the “object” to only text documents. User queries in IR are matched
against the documents’ representation stored in the database. Docu-
ments themselves often are not kept or stored directly in the IR sys-
tem. Rather, they are represented in the system by metadata. Typical
IR systems compute a numeric score on how well each document in
the database matches the query, and rank the objects according to this
value. The top-ranking documents from the system are then shown to

308
9.2. Semantic hashing with deep autoencoders for document 309

the user. The process may then be iterated if the user wishes to refine
the query.
Based partly on [236], common IR methods consist of several
categories:

• Boolean retrieval, where a document either matches a query or

does not.
• Algebraic approaches to retrieval, where models are used to rep-
resent documents and queries as vectors, matrices, or tuples. The
similarity of the query vector and document vector is represented
as a scalar value. This value can be used to produce a list of doc-
uments that are rank-ordered for a query. Common models and
methods include vector space model, topic-based vector space
model, extended Boolean model, and latent semantic analysis.
• Probabilistic approaches to retrieval, where the process of IR
is treated as a probabilistic inference. Similarities are computed
as probabilities that a document is relevant for a given query,
and the probability value is then used as the score in ranking
documents. Common models and methods include binary
independence model, probabilistic relevance model with the
BM25 relevance function, methods of inference with uncertainty,
probabilistic, language modeling, http://en.wikipedia.org/wiki/
Uncertain_inference and the technique of latent Dirichlet
allocation.
• Feature-based approaches to retrieval, where documents are
viewed as vectors of values of feature functions. Principled meth-
ods of “learning to rank” are devised to combine these features
into a single relevance score. Feature functions are arbitrary
functions of document and query, and as such Feature-based
approaches can easily incorporate almost any other retrieval
model as just yet another feature.

Deep learning applications to IR are rather recent. The approaches

in the literature so far belong mostly to the category of feature-based
approaches. The use of deep networks is mainly for extracting seman-
tically meaningful features for subsequent document ranking stages.
310 Selected Applications in Information Retrieval

We will review selected studies in the recent literature in the remain-

der of this section below.

9.2 Semantic hashing with deep autoencoders for document

indexing and retrieval

Here we discuss the “semantic hashing” approach for the application

of deep autoencoders to document indexing and retrieval as published
in [159, 314]. It is shown that the hidden variables in the final layer of
a DBN not only are easy to infer after using an approximation based
on feed-forward propagation, but they also give a better representation
of each document, based on the word-count features, than the widely
used latent semantic analysis and the traditional TF-IDF approach
for information retrieval. Using the compact code produced by deep
autoencoders, documents are mapped to memory addresses in such a
way that semantically similar text documents are located at nearby
addresses to facilitate rapid document retrieval. The mapping from a
word-count vector to its compact code is highly efficient, requiring only
a matrix multiplication and a subsequent sigmoid function evaluation
for each hidden layer in the encoder part of the network.
A deep generative model of DBN is exploited for the above purpose
as discussed in [165]. Briefly, the lowest layer of the DBN represents
the word-count vector of a document and the top layer represents a
learned binary code for that document. The top two layers of the DBN
form an undirected associative memory and the remaining layers form
a Bayesian (also called belief) network with directed, top-down connec-
tions. This DBN, composed of a set of stacked RBMs as we reviewed
in Section 5, produces a feed-forward “encoder” network that converts
word-count vectors to compact codes. By composing the RBMs in the
opposite order, a “decoder” network is constructed that maps com-
pact code vectors into reconstructed word-count vectors. Combining
the encoder and decoder, one obtains a deep autoencoder (subject to
further fine-tuning as discussed in Section 4) for document coding and
subsequent retrieval.
After the deep model is trained, the retrieval process starts with
mapping each query into a 128-bit binary code by performing a forward
9.3. DSSM for document retrieval 311

pass through the model with thresholding. Then the Hamming dis-
tance between the query binary code and all the documents’ 128-bit
binary codes, especially those of the “neighboring” documents defined
in the semantic space, are computed extremely efficiently. The effi-
ciency is accomplished by looking up the neighboring bit vectors in
the hash table. The same idea as discussed here for coding text docu-
ments for information retrieval has been explored for audio document
retrieval and speech feature coding problems with some initial explo-
ration reported in [100], discussed in Section 4 in detail.

9.3 Deep-structured semantic modeling (DSSM)

for document retrieval

Here we discuss the more advanced and recent approach to large-scale

document retrieval (Web search) based on a specialized deep architec-
ture, called deep-structured semantic model or deep semantic similarity
model (DSSM), as published in [172], and its convolutional version (C-
DSSM), as published in [328].
Modern search engines retrieve Web documents mainly by match-
ing keywords in documents with those in a search query. However, lex-
ical matching can be inaccurate due to the fact that a concept is often
expressed using different vocabularies and language styles in documents
and queries. Latent semantic models are able to map a query to its rel-
evant documents at the semantic level where lexical-matching often
fails [236]. These models address the language discrepancy between
Web documents and search queries by grouping different terms that
occur in a similar context into the same semantic cluster. Thus, a query
and a document, represented as two vectors in the lower-dimensional
semantic space, can still have a high similarity even if they do not share
any term. Probabilistic topic models such as probabilistic latent seman-
tic models and latent Dirichlet allocation models have been proposed
for semantic matching to partially overcome such difficulties. However,
the improvement on IR tasks has not been as significant as originally
expected because of two main factors: (1) most state-of-the-art latent
semantic models are based on linear projection, and thus are inadequate
in capturing effectively the complex semantic properties of documents;
312 Selected Applications in Information Retrieval

and (2) these models are often trained in an unsupervised manner using
an objective function that is only loosely coupled with the evaluation
metric for the retrieval task. In order to improve semantic matching for
IR, two lines of research have been conducted to extend the above latent
semantic models. The first is the semantic hashing approach reviewed
in Section 9.1 above in this section based on the use of deep autoen-
coders [165, 314]. While the hierarchical semantic structure embedded
in the query and the document can be extracted via deep learning,
the deep learning approach used for their models still adopts an unsu-
pervised learning method where the model parameters are optimized
for the re-construction of the documents rather than for differentiating
the relevant documents from the irrelevant ones for a given query. As
a result, the deep neural network models do not significantly outper-
form strong baseline IR models that are based on lexical matching. In
the second line of research, click-through data, which consists of a list
of queries and the corresponding clicked documents, is exploited for
semantic modeling so as to bridge the language discrepancy between
search queries and Web documents in recent studies [120, 124]. These
models are trained on click-through data using objectives that tailor to
the document ranking task. However, these click-through-based models
are still linear, suffering from the issue of expressiveness. As a result,
these models need to be combined with the keyword matching models
(such as BM25) in order to obtain a significantly better performance
than baselines.
The DSSM approach reported in [172] aims to combine the
strengths of the above two lines of work while overcoming their weak-
nesses. It uses the DNN architecture to capture complex semantic prop-
erties of the query and the document, and to rank a set of documents
for a given query. Briefly, a nonlinear projection is performed first to
map the query and the documents to a common semantic space. Then,
the relevance of each document given the query is calculated as the
cosine similarity between their vectors in that semantic space. The
DNNs are trained using the click-through data such that the condi-
tional likelihood of the clicked document given the query is maximized.
Different from the previous latent semantic models that are learned
in an unsupervised fashion, the DSSM is optimized directly for Web
9.3. DSSM for document retrieval 313

Figure 9.1: The DNN component of the DSSM architecture for computing semantic
features. The DNN uses multiple layers to map high-dimensional sparse text features,
for both Queries and Documents into low-dimensional dense features in a semantic
space. [after [172], @CIKM].

document ranking, and thus gives superior performance. Furthermore,

to deal with large vocabularies in Web search applications, a new word
hashing method is developed, through which the high-dimensional term
vectors of queries or documents are projected to low-dimensional letter
based n-gram vectors with little information loss.
Figure 9.1 illustrates the DNN part in the DSSM architecture. The
DNN is used to map high-dimensional sparse text features into low-
dimensional dense features in a semantic space. The first hidden layer,
with 30k units, accomplishes word hashing. The word-hashed features
are then projected through multiple layers of non-linear projections.
The final layer’s neural activities in this DNN form the feature in the
semantic space.
To show the computational steps in the various layers of the DNN
in Figure 9.1, we denote x as the input term vector, y as the output
vector, li , i = 1, . . . , N − 1, as the intermediate hidden layers, Wi as
the ith projection matrix, and bi as the ith bias vector, we have

l1 = W1 x,
li = f (Wi li−1 + bi ), i>1
y = f (WN lN −1 + bN ),
314 Selected Applications in Information Retrieval

where tanh function is used at the output layer and the hidden layers
li , i = 2, . . . , N − 1:
1 − e−2x
f (x) = .
1 + e−2x
The semantic relevance score between a query Q and a document D
can then be computed as the consine distance
Ty
yQ D
R(Q, D) = cosine(yQ , yD ) = ,
yQ yD 

where yQ and yD are the concept vectors of the query and the docu-
ment, respectively. In Web search, given the query, the documents can
be sorted by their semantic relevance scores.
Learning of the DNN weights Wi and bi shown in Figure 9.1 is an
important contribution of the study of [172]. Compared with the DNNs
used in speech recognition where the targets or labels of the training
data are readily available, the DNN in the DSSM does not have such
label information well defined. That is, rather than using the common
cross entropy or mean square errors as the training objective function,
IR-centric loss functions need to be developed in order to train the DNN
weights in the DSSM using the available data such as click-through logs.
The click-through logs consist of a list of queries and their clicked
documents. A query is typically more relevant to the documents that
are clicked on than those that are not. This weak supervision informa-
tion can be exploited to train the DSSM. More specifically, the weight
matrices in the DSSM, Wi , is learned to maximize the posterior prob-
ability of the clicked documents given the queries

exp(γR(Q, D))
P (D | Q) =

D  ∈D exp(γR(Q, D ))

defined on the semantic relevance score R(Q, D) between the Query (Q)
and the Document (D), where γ is a smoothing factor set empirically
on a held-out data set, and D denotes the set of candidate documents
to be ranked. Ideally, D should contain all possible documents, as in
the maximum mutual information training for speech recognition where
all possible negative candidates may be considered [147]. However in
9.3. DSSM for document retrieval 315

this case D is of Web scale and thus is intractable in practice. In the

implementation of DSSM learning described in [172], a subset of the
negative candidates are used, following the common practice adopted
in MCE (Minimum Classification Error) training in speech recognition
[52, 118, 417, 418]. In other words, for each query and clicked-document
pair, denoted by (QD+ ) where Q is a query and D+ is the clicked doc-
ument, the set of D is approximated by including D+ and only four
randomly selected unclicked documents, denoted by Dj− ; j = 1, . . . , 4}.
In the study reported in [172], no significant difference was found when
different sampling strategies were used to select the unclicked docu-
ments.
With the above simplification the DSSM parameters are estimated
to maximize the approximate likelihood of the clicked documents given
the queries across the training set

L(Λ) = log P (D+ | Q),

(Q,D + ,Dj− )

where Λ denotes the parameter set of the DNN weights {Wi } in the
DSSM. In Figure 9.2, we show the overall DSSM architecture that
contains several DNNs. All these DNNs share the same weights but take
different documents (one positive and several negatives) as inputs when
training the DSSM parameters. Details of the gradient computation
of this approximate loss function with respect to the DNN weights
tied across documents and queries can be found in [172] and are not
elaborated here.
Most recently, the DSSM described above has been extended to its
convolutional version, or C-DSSM [328]. In the C-DSSM, semantically
similar words within context are projected to vectors that are close
to each other in the contextual feature space through a convolutional
structure. The overall semantic meaning of a sentence is found to be
determined by a few key words in the sentence, and thus the C-DSSM
uses an additional max pooling layer to extract the most salient local
features to form a fixed-length global feature vector. The global feature
vector is then fed to the remaining nonlinear DNN layer(s) to map it
to a point in the shared semantic space.
316 Selected Applications in Information Retrieval

Figure 9.2: Architectural illustration of the DSSM for document retrieval (from
[170, 171]). All DNNs shown have shared weights. A set of n documents are shown
here to illustrate the random negative sampling discussed in the text for simplifying
the training procedure for the DSSM. [after [172], @CIKM].

The convolutional neural network component of the C-DSSM is

shown in Figure 9.3, where a window size of three is illustrated for
the convolutional layer. The overall C-DSSM architecture is similar
to the DSSM architecture shown in Figure 9.2 except that the fully-
connected DNNs are replaced by the convolutional neural networks
with locally-connected tied weights and additional max-pooling layers.
The model component shown in Figure 9.3 contains (1) a word hashing
layer to transform words into letter-tri-gram count vectors in the same
way as the DSSM; (2) a convolutional layer to extract local contextual
features for each context window; (3) a max-pooling layer to extract
and combine salient local contextual features to form a global feature
vector; and (4) a semantic layer to represent the high-level semantic
information of the input word sequence.
The main motivation for using the convolutional structure in the
C-DSSM is its ability to map a variable-length word sequence to a low-
dimensional vector in a latent semantic space. Unlike most previous
models that treat a query or a document as a bag of words, a query
or a document in the C-DSSM is viewed as a sequence of words with
contextual structures. By using the convolutional structure, local con-
textual information at the word n-gram level is modeled first. Then,
9.4. Use of deep stacking networks for information retrieval 317

Figure 9.3: The convolutional neural network component of the C-DSSM, with the
window size of three is illustrated for the convolutional layer. [after [328], @WWW].

salient local features in a word sequence are combined to form a global

feature vector. Finally, the high-level semantic information of the word
sequence is extracted to form a global vector representation. Like the
DSSM just described, the C-DSSM is also trained on click-through data
by maximizing the conditional likelihood of the clicked documents given
a query using the back-propagation algorithm.

9.4 Use of deep stacking networks for information retrieval

In parallel with the IR studies reviewed above, the deep stacking net-
work (DSN) discussed in Section 6 has also been explored recently
for IR with insightful results [88]. The experimental results suggest
that the classification error rate using the binary decision of “relevant”
versus “non-relevant” from the DSN, which is closely correlated with
the DSN training objective, is also generally correlated well with the
NDCG (normalized discounted cumulative gain) as the most common
318 Selected Applications in Information Retrieval

IR quality measure. The exception is found in the region of high IR

quality.
As described in Section 6, the simplicity of the DSN’s training objec-
tive, the mean square error (MSE), drastically facilitates its success-
ful applications to image recognition, speech recognition, and speech
understanding. The MSE objective and classification error rate have
been shown to be well correlated in these speech or image applications.
For information retrieval (IR) applications, however, the inconsistency
between the MSE objective and the desired objective (e.g., NDCG)
is much greater than that for the above classification-focused applica-
tions. For example, the NDCG as a desirable IR objective function is
a highly non-smooth function of the parameters to be learned, with a
very different nature from the nonlinear relationship between MSE and
classification error rate. Thus, it is of interest to understand to what
extent the NDCG is reasonably well correlated with classification rate
or MSE where the relevance level in IR is used as the DSN prediction
target. Further, can the advantage of learning simplicity in the DSN
be applied to improve IR quality measures such as the NDCG? Our
experimental results presented in [88] provide largely positive answers
to both of the above questions. In addition, special care that need to
be taken in implementing DSN learning algorithms when moving from
classification to IR applications are addressed.
The IR task in the experiments of [88] is the sponsored search
related to ad placement. In addition to the organic web search
results, commercial search engines also provide supplementary spon-
sored results in response to the user’s query. The sponsored search
results are selected from a database pooled by advertisers who bid to
have their ads displayed on the search result pages. Given an input
query, the search engine will retrieve relevant ads from the database,
rank them, and display them at the proper place on the search result
page; e.g., at the top or right hand side of the web search results. Find-
ing relevant ads to a query is quite similar to common web search. For
instance, although the documents come from a constrained database,
the task resembles typical search ranking that targets on predicting
document relevance to the input query. The experiments conducted for
9.4. Use of deep stacking networks for information retrieval 319

this task are the first with the use of deep learning techniques (based
on the DSN architecture) on the ad-related IR problem. The prelimi-
nary results from the experiments are the close correlation between the
MSE as the DSN training objective with the NDCG as the IR quality
measure over a wide NDCG range.
 10
Selected Applications in Object Recognition
and Computer Vision

Over the past two years or so, tremendous progress has been made in
applying deep learning techniques to computer vision, especially in the
field of object recognition. The success of deep learning in this area
is now commonly accepted by the computer vision community. It is
the second area in which the application of deep learning techniques
is successful, following the speech recognition area as we reviewed and
analyzed in Sections 2 and 7.
Excellent surveys on the recent progress of deep learning for
computer vision are available in the NIPS-2013 tutorial (https://
nips.cc/Conferences/2013/Program/event.php?ID=4170 with video
recording at http://research.microsoft.com/apps/video/default.aspx?
id=206976&l=i) and slides at http://cs.nyu.edu/∼fergus/presentations/
nips2013_final.pdf, and also in the CVPR-2012 tutorial (http://cs.nyu.
edu/∼fergus/tutorials/deep_learning_cvpr12). The reviews provided
in this section below are based partly on these tutorials, in connection
with the earlier deep learning material in this monograph. Another
excellent source which this section draws from is the most recent Ph.D.
thesis on the topic of deep learning for computer vision [434].

320
10.1. Unsupervised or generative feature learning 321

Over many years, object recognition in computer vision has been

relying on hand-designed features such as SIFT (scale invariant fea-
ture transform) and HOG (histogram of oriented gradients), akin to
the reliance of speech recognition on hand-designed features such as
MFCC and PLP. However, features like SIFT and HOG only capture
low-level edge information. The design of features to effectively capture
mid-level information such as edge intersections or high-level represen-
tation such as object parts becomes much more difficult. Deep learning
aims to overcome such challenges by automatically learning hierarchies
of visual features in both unsupervised and supervised manners directly
from data. The review below categorizes the many deep learning meth-
ods applied to computer vision into two classes: (1) unsupervised fea-
ture learning where the deep learning is used to extract features only,
which may be subsequently fed to relatively simple machine learning
algorithm for classification or other tasks; and (2) supervised learning
methods where end-to-end learning is adopted to jointly optimize fea-
ture extractor and classifier components of the full system when large
amounts of labeled training data are available.

10.1 Unsupervised or generative feature learning

When labeled data are relatively scarce, unsupervised learning algo-

rithms have been shown to learn useful visual feature hierarchies. In
fact, prior to the demonstration of remarkable successes of CNN archi-
tectures with supervised learning in the 2012 ImageNet competition,
much of the work in applying deep learning methods to computer
vision had been on unsupervised feature learning. The original unsuper-
vised deep autoencoder that exploits DBN pre-training was developed
and demonstrated by Hinton and Salakhutdinov [164] with success on
the image recognition and dimensionality reduction (coding) tasks of
MNIST with only 60,000 samples in the training set; see details of this
task in http://yann.lecun.com/exdb/mnist/ and an analysis in [78].
It is interesting to note that the gain of coding efficiency using the DBN-
based autoencoder on the image data over the conventional method of
principal component analysis as demonstrated in [164] is very similar to
322 Selected Applications in Object Recognition and Computer Vision

the gain reported in [100] and described in Section 4 of this monograph

on the speech data over the traditional technique of vector quantiza-
tion. Also, Nair and Hinton [265] developed a modified DBN where the
top-layer model uses a third-order Boltzmann machine. This type of
DBN is applied to the NORB database — a three-dimensional object
recognition task. An error rate close to the best published result on this
task is reported. In particular, it is shown that the DBN substantially
outperforms shallow models such as SVMs. In [358], two strategies to
improve the robustness of the DBN are developed. First, sparse connec-
tions in the first layer of the DBN are used as a way to regularize the
model. Second, a probabilistic de-noising algorithm is developed. Both
techniques are shown to be effective in improving robustness against
occlusion and random noise in a noisy image recognition task. DBNs
have also been successfully applied to create compact but meaning-
ful representations of images [360] for retrieval purposes. On this large
collection image retrieval task, deep learning approaches also produced
strong results. Further, the use of a temporally conditional DBN for
video sequence and human motion synthesis were reported in [361]. The
conditional RBM and DBN make the RBM and DBN weights associ-
ated with a fixed time window conditioned on the data from previous
time steps. The computational tool offered in this type of temporal
DBN and the related recurrent networks may provide the opportunity
to improve the DBN–HMMs towards efficient integration of temporal-
centric human speech production mechanisms into DBN-based speech
production model.
Deep learning methods have a rich family, including hierarchical
probabilistic and generative models (neural networks or otherwise).
One most recent example of this type developed and applied to facial
expression datasets is the stochastic feed-forward neural networks that
can be learned efficiently and that can induce a rich multiple-mode
distribution in the output space not possible with the standard, deter-
ministic neural networks [359]. In Figure 10.1, we show the architecture
of a typical stochastic feed-forward neural network with four hidden
layers with mixed deterministic and stochastic neurons (left) used to
model multi-mode distributions illustrated on the right. The stochastic
network here is a deep, directed graphical model, where the generation
10.1. Unsupervised or generative feature learning 323

Figure 10.1: Left: A typical architecture of the stochastic feed-forward neural

network with four hidden layers. Right: Illustration of how the network can produce
a distribution with two distinct modes and use them to represent two or more
different facial expressions y given a neutral face x. [after [359], @NIPS].

process starts from input x, a neural face, and generates the output
y, the facial expression. In face expression classification experiments,
the learned unsupervised hidden features generated from this stochas-
tic network are appended to the image pixels and helped to obtain
superior accuracy to the baseline classifier based on the conditional
RBM/DBN [361].
Perhaps the most notable work in the category of unsupervised deep
feature learning for computer vision (prior to the recent surge of the
work on CNNs) is that of [209], a nine-layer locally connected sparse
autoencoder with pooling and local contrast normalization. The model
has one billion connections, trained on the dataset with 10 million
images downloaded from the Internet. The unsupervised feature learn-
ing methods allow the system to train a face detector without having to
label images as containing a face or not. And the control experiments
show that this feature detector is robust not only to translation but
also to scaling and out-of-plane rotation.
Another set of popular studies on unsupervised deep feature learn-
ing for computer vision are based on deep sparse coding models [226].
This type of deep models produced state-of-the-art accuracy results on
the ImageNet object recognition tasks prior to the rise of the CNN
architectures armed with supervised learning to perform joint feature
learning and classification, which we turn to now.
324 Selected Applications in Object Recognition and Computer Vision

10.2 Supervised feature learning and classification

The origin of the applications of deep learning to object recognition

tasks can be traced to the convolutional neural networks (CNNs)
in the early 90s; see a comprehensive overview in [212]. The CNN-
based architectures in the supervised learning mode have captured
intense interest in computer vision since October 2012 shortly after
the ImageNet competition results were released (http://www.image-
net.org/challenges/LSVRC/2012/). This is mainly due to the huge
recognition accuracy gain over competing approaches when large
amounts of labeled data are available to efficiently train large CNNs
using GPU-like high-performance computing platforms. Just like DNN-
based deep learning methods have outperformed previous state-of-
the-art approaches in speech recognition in a series of benchmark
tasks including phone recognition, large-vocabulary speech recognition,
noise-robust speech recognition, and multi-lingual speech recognition,
CNN-based deep learning methods have demonstrated the same in a
set of computer vision benchmark tasks including category-level object
recognition, object detection, and semantic segmentation.
The basic architecture of the CNN described in [212] is shown in
Figure 10.1. To incorporate the relative invariance of the spatial rela-
tionship in typical image pixels with respect to the location, the CNN
uses a convolutional layer with local receptive fields and with tied fil-
ter weights, much like 2-dimensional FIR filters in image processing.
The output of the FIR filters is then passed through a nonlinear acti-
vation function to create activation maps, followed by another non-
linear pooling (labeled as “subsampling” in Figure 10.2) layer that
reduces the data rate while providing invariance to slightly differ-
ent input images. The output of the pooling layer is fed to a few
fully connected layers as in the DNN discussed in earlier chapters.
The whole architecture above is also called the deep CNN in the
literature.
Deep models with convolution structure such as CNNs have been
found effective and have been in use in computer vision and image
recognition since 90s [57, 185, 192, 198, 212]. The most notable advance
was achieved in the 2012 ImageNet LSVRC competition, in which
10.2. Supervised feature learning and classification 325

Figure 10.2: The original convolutional neural network that is composed of mul-
tiple alternating convolution and pooling layers followed by fully connected layers.
[after [212], @IEEE].

the task is to train a model with 1.2 million high-resolution images

to classify unseen images to one of the 1000 different image classes.
On the test set consisting of 150k images, the deep CNN approach
described in [198] achieved the error rates considerably lower than the
previous state-of-the-art. Very large deep-CNNs are used, consisting of
60 million weights, and 650,000 neurons, and five convolutional layers
together with max-pooling layers. Additional two fully-connected layers
as in the DNN described previously are used on top of the CNN layers.
Although all the above structures were developed separately in earlier
work, their best combination accounted for major part of the success.
See the overall architecture of the deep CNN system in Figure 10.3. Two
additional factors contribute to the final success. The first is a powerful
regularization technique called “dropout”; see details in [166] and a
series of further analysis and improvement in [10, 13, 240, 381, 385]. In
particular, Warde-Farley et al. [385] analyzed the disentangling effects
of dropout and showed that it helps because different members of the
bag share parameters. Applications of the same “dropout” techniques
are also successful for some speech recognition tasks [65, 81]. The
second factor is the use of non-saturating neurons or rectified linear
units (ReLU) that compute f (x) = max(x, 0), which significantly
speeds up the overall training process especially with efficient GPU
implementation. This deep-CNN system achieved a winning top-5 test
error rate of 15.3% using extra training data from ImageNet Fall 2011
release, or 16.4% using only supplied training data in ImageNet-2012,
326 Selected Applications in Object Recognition and Computer Vision

Figure 10.3: The architecture of the deep-CNN system which won the 2012 Ima-
geNet competition by a large margin over the second-best system and the state of
the art by 2012. [after [198], @NIPS].

significantly lower than 26.2% achieved by the second-best system

which combines scores from many classifiers using a set of hand-
crafted features such as SIFT and Fisher vectors. See details in http://
www.image-net.org/challenges/LSVRC/2012/oxford_vgg.pdf about
the best competing method. It is noted, however, that the Fisher-
vector-encoding approach has recently been extended by Simonyan
et al. [329] via stacking in multiple layers to form deep Fisher net-
works, which achieve competitive results with deep CNNs at a smaller
computational learning cost.
The state of the art performance demonstrated in [198] using the
deep-CNN approach is further improved by another significant mar-
gin during 2013, using a similar approach but with bigger models
and larger amounts of training data. A summary of top-5 test error
rates from 11 top-performing teams participating in the 2013 Ima-
geNet ILSVRC competition is shown in Figure 10.4, with the best
result of the 2012 competition shown to the right most as the baseline.
Here we see rapid error reduction on the same task from the lowest
pre-2012 error rate of 26.2% (non-neural networks) to 15.3% in 2012
and further to 11.2% in 2013, both achieved with deep-CNN technol-
ogy. It is also interesting to observe that all major entries in the 2013
ImageNet ILSVRC competition is based on deep learning approaches.
For example, the Adobe system shown in Figure 10.4 is based on the
deep-CNN reported in [198] including the use of dropout. The network
10.2. Supervised feature learning and classification 327

Figure 10.4: Summary results of ImageNet Large Scale Visual Recognition

Challenge 2013 (ILSVRC2013), representing the state-of-the-are performance of
object recognition systems. Data source: http://www.image-net.org/challenges/
LSVRC/2013/results.php.

architecture is modified to include more filters and connections. At test

time, image saliency is used to obtain 9 crops from original images,
which are combined with the standard five multiview crops. The NUS
system uses a non-parametric, adaptive method to combine the out-
puts from multiple shallow and deep experts, including deep-CNN,
kernel, and GMM methods. The VGG system is described in [329]
and uses a combination of the deep Fisher vector network and the
deep-CNN. The ZF system is based on a combination of a large CNN
with a range of different architectures. The choice of architectures was
assisted by visualization of model features using a deconvolutional net-
work as described by Zeiler et al. [437], Zeiler and Fergus [435, 436],
and Zeiler ([434]). The CognitiveVision system uses an image classifi-
cation scheme based on a DNN architecture. The method is inspired
by cognitive psychophysics about how the human vision system first
learns to classify the basic-level categories and then learns to clas-
sify categories at the subordinate level for fine-grained object recogni-
tion. Finally, the best-performing system called Clarifai in Figure 10.4
is based on a large and deep CNN with dropout regularization. It
328 Selected Applications in Object Recognition and Computer Vision

augments the amount of training data by down-sampling images to

256 pixels. The system contains a total of 65M parameters. Multiple
such models were averaged together to further boost performance. The
main novelty is to use the visualization technique based on the deconvo-
lutional networks as described in [434, 437] to identify what makes the
deep model perform well, based on which a powerful deep architecture
was chosen. See more details of these systems in http://www.image-
net.org/challenges/LSVRC/2013/results.php.
While the deep CNN has demonstrated remarkable classification
performance on object recognition tasks, there has been no clear under-
standing of why they perform so well until recently. Zeiler and Fergus
[435, 436] conducted research to address just this issue, and then used
the gained understanding to further improve the CNN systems, which
yielded excellent performance as shown in Figure 10.4 with labels “ZF”
and “Clarifai.” A novel visualization technique is developed that gives
insight into the function of intermediate feature layers of the deep CNN.
The technique also sheds light onto the operation of the full network
acting as a classifier. The visualization technique is based on a decon-
volutional network, which maps the neural activities in intermediate
layers of the original convolutional network back to the input pixel
space. This allows the researchers to examine what input pattern orig-
inally caused a given activation in the feature maps. Figure 10.5 (the
top portion) illustrates how a deconvolutional network is attached to
each of its layers, thereby providing a closed loop back to image pixels
as the input to the original CNN. The information flow in this closed
loop is as follows. First, an input image is presented to the deep CNN in
a feed-forward manner so that the features at all layers are computed.
To examine a given CNN activation, all other activations in the layer
are set to zero and the feature maps are passed as input to the attached
deconvolutional network’s layer. Then, successive operations, opposite
to the feed-forward computation in the CNN, are carried out including
unpooling, rectifying, and filtering. This allows the reconstruction of
the activity in the layer beneath that gave rise to the chosen activa-
tion. These operations are repeated until input layer is reached. During
unpooling, non-invertibility of the max pooling operation in the CNN is
10.2. Supervised feature learning and classification 329

Figure 10.5: The top portion shows how a deconvolutional network’s layer (left)
is attached to a corresponding CNN’s layer (right). The d econvolutional network
reconstructs an approximate version of the CNN features from the layer below. The
bottom portion is an illustration of the unpooling operation in the deconvolutional
network, where “Switches” are used to record the location of the local max in each
pooling region during pooling in the CNN. [after [436], @arXiv].

resolved by an approximate inverse, where the locations of the maxima

within each pooling region are recorded in a set of “switch” variables.
These switches are used to place the reconstructions from the layer
above into appropriate locations, preserving the structure of the stim-
ulus. This procedure is shown at the bottom portion of Figure 10.5.
330 Selected Applications in Object Recognition and Computer Vision

In addition to the deep-CNN architecture described above, the DNN

architecture has also been shown to be highly successful in a number
of computer vision tasks [54, 55, 56, 57]. We have not found in the
literature on direct comparisons among the CNN, DNN, and other
related architectures on the identical tasks.
Finally, the most recent study on supervised learning for computer
vision shows that the deep CNN architecture is not only successful for
object/image classification discussed earlier in this section but also suc-
cessful for objection detection in the whole images [128]. The detection
task is substantially more complex than the classification task.
As a brief summary of this chapter, deep learning has made huge
inroads into computer vision, soon after its success in speech recogni-
tion discussed in Section 7. So far, it is the supervised learning paradigm
based on the deep CNN architecture and the related classification tech-
niques that are making the greatest impact, showcased by the ImageNet
competition results from 2012 and 2013. These methods can be used
for not only object recognition but also many other computer vision
tasks. There has been some debate as to the reasons for the success of
these CNN-based deep learning methods, and about their limitations.
Many questions are still open as to how these methods can be tai-
lored to certain computer vision applications and how to scale up the
models and training data. Finally, we discussed a number of studies on
unsupervised and generative approaches of deep learning to computer
vision and image modeling problems in the earlier part of this chapter.
Their performance has not been competitive with the supervised learn-
ing approach on object recognition tasks with ample training data. To
achieve long term and ultimate success in computer vision, it is likely
that unsupervised learning will be needed. To this end, many open
problems in unsupervised feature learning and deep learning need to
be addressed and much more research need to be carried out.
 11
Selected Applications in Multimodal
and Multi-task Learning

Multi-task learning is a machine learning approach that learns to solve

several related problems at the same time, using a shared represen-
tation. It can be regarded as one of the two major classes of transfer
learning or learning with knowledge transfer, which focuses on general-
izations across distributions, domains, or tasks. The other major class
of transfer learning is adaptive learning, where knowledge transfer is
carried out in a sequential manner, typically from a source task to a
target task [95]. Multi-modal learning is a closely related concept to
multi-task learning, where the learning domains or “tasks” cut across
several modalities for human–computer interactions or other applica-
tions embracing a mixture of textual, audio/speech, touch, and visual
information sources.
The essence of deep learning is to automate the process of dis-
covering effective features or representations for any machine learn-
ing task, including automatically transferring knowledge from one task
to another concurrently. Multi-task learning is often applied to con-
ditions where no or very little training data are available for the tar-
get task domain, and hence is sometimes called zero-shot or one-shot
learning. It is evident that difficult multi-task leaning naturally fits the
paradigm of deep learning or representation learning where the shared

331
332 Selected Applications in Multimodal and Multi-task Learning

representations and statistical strengths across tasks (e.g., those involv-

ing separate modalities of audio, image, touch, and text) is expected
to greatly facilitate many machine learning scenarios under low- or
zero-resource conditions. Before deep learning methods were adopted,
there had been numerous efforts in multi-modal and multi-task learn-
ing. For example, a prototype called MiPad for multi-modal interac-
tions involving capturing, leaning, coordinating, and rendering a mix
of speech, touch, and visual information was developed and reported
in [175, 103]. And in [354, 443], mixed sources of information from
multiple-sensory microphones with separate bone-conductive and air-
born paths were exploited to de-noise speech. These early studies all
used shallow models and learning methods and achieved worse than
desired performance. With the advent of deep learning, it is hopeful
that the difficult multi-modal learning problems can be solved with
eventual success to enable a wide range of practical applications. In
this chapter, we will review selected applications in this area, orga-
nized according to different combinations of more than one modalities
or learning tasks. Much of the work reviewed here is on-going research,
and readers should expect follow-up publications in the future.

11.1 Multi-modalities: Text and image

The underlying mechanism for potential effectiveness of multi-modal

learning involving text and image is the common semantics associated
with the text and image. The relationship between the text and image
may come, for example, from the text annotations of an image (as the
training data for a multi-modal learning system). If the related text
and image share the same representation in a common semantic space,
the system can generalize to the unseen situation where either text
or image is unavailable. It can thus be naturally used for zero-shot
learning for image or text. In other words, multi-modality learning can
use text information to help image/visual recognition, and vice versa.
Exploiting text information for image/visual recognition constitutes
most of the work done in this space, which we review in this section
below.
11.1. Multi-modalities: Text and image 333

The deep architecture, called DeViSE (deep visual-semantic embed-

ding) and developed by Frome et al. [117], is a typical example of the
multi-modal learning where text information is used to improve the
image recognition system, especially for performing zero-shot learning.
Image recognition systems are often limited in their ability to scale
to large number of object categories, due in part to the increasing
difficulty of acquiring sufficient training data with text labels as the
number of image categories grows. The multi-modal DeViSE system
is aimed to leverage text data to train the image models. The joint
model is trained to identify image classes using both labeled image
data and the semantic information learned from unannotated text. An
illustration of the DeViSE architecture is shown in the center portion
of Figure 10.1. It is initialized with the parameters pre-trained at the
lower layers of two models: the deep-CNN for image classification in
the left portion of the figure and the text embedding model in the
right portion of the figure. The part of the deep CNN, labeled “core
visual model” in Figure 10.1, is further learned to predict the target
word-embedding vector using a projection layer labeled “transforma-
tion” and using a similarity metric. The loss function used in training
adopts a combination of dot-product similarity and max-margin, hinge
rank loss. The former is the un-normalized version of the cosine loss
function used for training the DSSM model in [170] as described in
Section 9.3. The latter is similar to the earlier joint image-text model
called WSABIE (web scale annotation by image embedding developed
by Weston et al. [388, 389]. The results show that the information pro-
vided by text improves zero-shot image predictions, achieving good hit
rates (close to 15%) across thousands of the labels never seen by the
image model.
The earlier WSABIE system as described in [388, 389] adopted
a shallow architecture and trained a joint embedding model of both
images and labels. Rather than using deep architectures to derive the
highly nonlinear image (as well as text-embedding) feature vectors as in
DeViSE, the WSABIE uses simple image features and a linear mapping
to arrive at the joint embedding space. Further, it uses an embedding
vector for each possible label. Thus, unlike DeViSE, WSABIE could
not generalize to new classes.
334 Selected Applications in Multimodal and Multi-task Learning

Figure 11.1: Illustration of the multi-modal DeViSE architecture. The left portion
is an image recognition neural network with a softmax output layer. The right por-
tion is a skip-gram text model providing word embedding vectors; see Section 8.2
and Figure 8.3 for details. The center is the joint deep image-text model of DeViSE,
with the two Siamese branches initialized by the image and word embedding mod-
els below the softmax layers. The layer labeled “transformation” is responsible for
mapping the outputs of the image (left) and text (right) branches into the same
semantic space. [after [117], @NIPS].

It is also interesting to compare the DeViSE architecture of

Figure 11.1 with the DSSM architecture of Figure 9.2 in Section 9.
The branches of “Query” and “Documents” in DSSM are analogous to
the branches of “image” and “text-label” in DeViSE. Both DeViSE and
DSSM use the objective function related to cosine distance between
two vectors for training the network weights in an end-to-end fash-
ion. One key difference, however, is that the two sets of inputs to the
DSSM are both text (i.e., “Query” and “Documents” designed for IR),
and thus mapping “Query” and “Documents” to the same semantic
space is conceptually more straightforward compared with the need
in DeViSE for mapping from one modality (image) to another (text).
Another key difference is that the generalization ability of DeViSE to
unseen image classes comes from computing text embedding vectors
for many unsupervised text sources (i.e., with no image counterparts)
that would cover the text labels corresponding to the unseen classes.
The generalization ability of the DSSM over unseen words, however,
is derived from a special coding scheme for words in terms of their
constituent letters.
The DeViSE architecture has inspired a more recent method,
which maps images into the semantic embedding space via convex
11.1. Multi-modalities: Text and image 335

combination of embedding vectors for the text label and the image
classes [270]. Here is the main difference. DeViSE replaces the last,
softmax layer of a CNN image classifier with a linear transformation
layer. The new transformation layer is then trained together with the
lower layers of the CNN. The method in [270] is much simpler — keep-
ing the softmax layer of the CNN while not training the CNN. For a
test image, the CNN first produces top N-best candidates. Then, the
convex combination of the corresponding N embedding vectors in the
semantic space is computed. This gives a deterministic transformation
from the outputs of the softmax classifier into the embedding space.
This simple multi-modal learning method is shown to work very well
on the ImageNet zero-shot learning task.
Another thread of studies separate from but related to the above
work on multi-modal learning involving text and image have cen-
tered on the use of multi-modal embeddings, where data from multiple
sources with separate modalities of text and image are projected into
the same vector space. For example, Socher and Fei-Fei [341] project
words and images into the same space using kernelized canonical cor-
relation analysis. Socher et al. [342] map images to single-word vectors
so that the constructed multi-modal system can classify images with-
out seeing any examples of the class, i.e., zero-shot learning similar
to the capability of DeViSE. The most recent work by Socher et al.
[343] extends their earlier work from single-word embeddings to those
of phrases and full-length sentences. The mechanism for mapping sen-
tences instead of the earlier single words into the multi-modal embed-
ding space is derived from the power of the recursive neural network
described in Socher et al. [347] as summarized in Section 8.2, and its
extension with dependency tree.
In addition to mapping text to image (or vice versa) into the same
vector space or to creating the joint image/text embedding space,
multi-modal learning for text and image can also be cast in the frame-
work of language models. In [196], a model of natural language is made
conditioned on other modalities such as image as the focus of the
study. This type of multi-modal language model is used to (1) retrieve
images given complex description queries, (2) retrieve phrase descrip-
tions given image queries, and (3) generate text conditioned on images.
336 Selected Applications in Multimodal and Multi-task Learning

Figure 11.2: Illustration of the multi-modal DeViSE architecture. The left portion
is an image recognition neural network with a softmax output layer. The right por-
tion is a skip-gram text model providing word embedding vectors; see Section 8.2
and Figure 8.3 for details. The center is the joint deep image-text model of DeViSE,
with the two Siamese branches initialized by the image and word embedding mod-
els below the softmax layers. The layer labeled “transformation” is responsible for
mapping the outputs of the image (left) and text (right) branches into the same
semantic space. [after [196], @NIPS].

Word representations and image features are jointly learned by train-

ing the multi-modal language model together with a convolutional net-
work. An illustration of the multi-modal language model is shown in
Figure 11.2.

11.2 Multi-modalities: Speech and image

Ngiam et al. [268, 269] propose and evaluate an application of

deep networks to learn features over audio/speech and image/video
modalities. They demonstrate cross-modality feature learning, where
better features for one modality (e.g., image) is learned when multiple
modalities (e.g., speech and image) are present at feature learning time.
A bi-modal deep autoencoder architecture for separate audio/speech
and video/image input channels are shown in Figure 11.3. The essence
of this architecture is to use a shared, middle layer to represent both
types of modalities. This is a straightforward generalization from
the single-modal deep autoencoder for speech shown in Figure 4.1 of
Section 4 to bi-modal counterpart. The authors further show how to
11.2. Multi-modalities: Speech and image 337

Figure 11.3: The architecture of a deep denoising autoencoder for multi-modal

audio/speech and visual features. [after [269], @ICML].

learn a shared audio and video representation, and evaluate it on a

fixed task, where the classifier is trained with audio-only data but
tested with video-only data and vice versa. The work concludes that
deep learning architectures are generally effective in learning multi-
modal features from unlabeled data and in improving single modality
features through cross modality information transfer. One exception
is the cross-modality setting using the CUAVE dataset. The results
presented in [269, 268] show that learning video features with both
video and audio outperforms that with only video data. However, the
same paper also shows that a model of [278] in which a sophisticated
signal processing technique for extracting visual features, together
with the uncertainty-compensation method developed originally from
robust speech recognition [104], gives the best classification accuracy
in the cross-modal learning task, beating the features derived from the
generative deep architecture designed for this task.
While the deep generative architecture for multimodal learning
described in [268, 269] is based on non-probabilistic autoencoder neural
338 Selected Applications in Multimodal and Multi-task Learning

nets, a probabilistic version based on deep Boltzmann machine (DBM)

has appeared more recently for the same multimodal application. In
[348], a DBM is used to extract a unified representation integrat-
ing separate modalities, useful for both classification and information
retrieval tasks. Rather than using the “bottleneck” layers in the deep
autoencoder to represent multimodal inputs, here a probability den-
sity is defined on the joint space of multimodal inputs, and states of
suitably defined latent variables are used for the representation. The
advantage of this probabilistic formulation, possibly lacking in the tra-
ditional deep autoencoder, is that the missing modality’s information
can be filled in naturally by sampling from its conditional distribution.
More recent work on autoencoders [22, 30] shows the capability of gen-
eralized denoising autoencoders in carrying out sampling, thus they
may overcome the earlier problem of filling-in the missing modality’s
information. For the bi-modal data consisting of image and text, the
multimodal DBM was shown to slightly outperform the traditional ver-
sion of the deep multimodal autoencoder as well as multimodal DBN
in classification and information retrieval tasks. No results on the com-
parisons with the generalized version of deep autoencoders has been
reported but may appear soon.
The several architectures discussed so far in this chapter for multi-
modal processing and learning can be regarded as special cases of
more general multi-task learning and transfer learning [22, 47]. Trans-
fer learning, encompassing both adaptive and multi-task learning, refers
to the ability of a learning architecture and technique to exploit com-
mon hidden explanatory factors among different learning tasks. Such
exploitation permits sharing of aspects of diverse types of input data
sets, thus allowing the possibility of transferring knowledge across seem-
ingly different learning tasks. As argued in [22], the learning archi-
tecture shown in Figure 11.4 and the associated learning algorithms
have an advantage for such tasks because they learn representations
that capture underlying factors, a subset of which may be relevant
for each particular task. We will discuss a number of such multi-task
learning applications in the remainder of this chapter that are confined
with a single modality of speech, natural language processing, or image
domain.
11.3 ML within the speech, NLP or image domain 339

Figure 11.4: A DNN architecture for multitask learning that is aimed to dis-
cover hidden explanatory factors shared among three tasks A, B, and C. [after [22],
@IEEE].

11.3 Multi-task learning within the speech, NLP or image

domain

Within the speech domain, one most interesting application of multi-

task learning is multi-lingual or cross-lingual speech recognition, where
speech recognition for different languages is considered as different
tasks. Various approaches have been taken to attack this rather chal-
lenging acoustic modeling problem for speech recognition, where the
difficulty lies in the lack of transcribed speech data due to economic
considerations in developing speech recognition systems for all lan-
guages in the world. Cross-language data sharing and data weighing
are common and useful approaches for the GMM–HMM system [225].
Another successful approach for the GMM–HMM is to map pronunci-
ation units across languages either via knowledge-based or data-driven
methods [420]. But they are much inferior to the DNN–HMM approach
which we now summarize.
In recent papers of [94, 170] and [150], two research groups inde-
pendently developed closely related DNN architectures with multi-task
learning capabilities for multilingual speech recognition. See Figure 11.5
for an illustration of this type of architecture. The idea behind these
architectures is that the hidden layers in the DNN, when learned
340 Selected Applications in Multimodal and Multi-task Learning

Figure 11.5: A DNN architecture for multilingual speech recognition. [after [170],
@IEEE].

appropriately, serve as increasingly complex feature transformations

sharing common hidden factors across the acoustic data in different
languages. The final softmax layer representing a log-linear classifier
makes use of the most abstract feature vectors represented in the top-
most hidden layer. While the log-linear classifier is necessarily sepa-
rate for different languages, the feature transformations can be shared
across languages. Excellent multilingual speech recognition results are
reported, far exceeding the earlier results using the GMM–HMM based
approaches [225, 420]. The implication of this set of work is signif-
icant and far reaching. It points to the possibility of quickly build-
ing a high-performance DNN-based system for a new language from
an existing multilingual DNN. This huge benefit would require only a
small amount of training data from the target language, although hav-
ing more data would further improve the performance. This multitask
learning approach can reduce the need for the unsupervised pre-training
stage, and can train the DNN with much fewer epochs. Extension
of this set of work would be to efficiently build a language-universal
speech recognition system. Such a system cannot only recognize many
languages and improve the accuracy for each individual language, but
11.3. ML within the speech, NLP or image domain 341

Figure 11.6: A DNN architecture for speech recognition trained with mixed-
bandwidth acoustic data with 16-kHz and 8-kHz sampling rates; [after [221],
@IEEE].

also expand the languages supported by simply stacking softmax layers

on the DNN for new languages.
A closely related DNN architecture, as shown in Figure 11.6, with
multitask learning capabilities was also recently applied to another
acoustic modeling problem — learning joint representations for two
separate sets of acoustic data [94, 221]. The set that consists of the
speech data with 16 kHz sampling rate is of wideband and high qual-
ity, which is often collected from increasingly popular smart phones
under the voice search scenario. Another, narrowband data set has a
lower sampling rate of 8kHz, often collected using the telephony speech
recognition systems.
As a final example of multi-task learning within the speech domain,
let us consider phone recognition and word recognition as separate
“tasks.” That is, phone recognition results are used not for producing
text outputs but for language-type identification or for spoken doc-
ument retrieval. Then, the use of pronunciation dictionary in almost
all speech systems can be considered as multi-task learning that share
the tasks of phone recognition and word recognition. More advanced
frameworks in speech recognition have pushed this direction further
342 Selected Applications in Multimodal and Multi-task Learning

by advocating the use of even finer units of speech than phones to

bridge the raw acoustic information of speech to semantic content of
speech via a hierarchy of linguistic structure. These atomic speech units
include “speech attributes” in the detection-based and knowledge-rich
modeling framework for speech recognition, whose accuracy has been
significantly boosted recently by the use of deep learning methods
[332, 330, 427].
Within the natural language processing domain, the best known
example of multi-task learning is the comprehensive studies reported
in [62, 63], where a range of separate “tasks” of part-of-speech tag-
ging, chunking, named entity tagging, semantic role identification, and
similar-word identification in natural language processing are attacked
using a common representation of words and a unified deep learning
approach. A summary of these studies can be found in Section 8.2.
Finally, within the domain of image/vision as a single modality,
deep learning has also been found effective in multi-task learning. Sri-
vastava and Salakhutdinov [349] present a multi-task learning approach
based on hierarchical Bayesian priors in a DNN system applied to var-
ious image classification data sets. The priors are combined with a
DNN, which improves discriminative learning by encouraging infor-
mation sharing among tasks and by discovering similar classes among
which knowledge is transferred. More specifically, methods are devel-
oped to jointly learn to classify images and a hierarchy of classes, such
that “poor classes,” for which there are relatively few training examples,
can benefit from similar “rich classes,” for which more training exam-
ples are available. This work can be considered as an excellent instance
of learning output representations, in addition to learning input rep-
resentation of the DNN as the focus of nearly all deep learning work
reported in the literature.
As another example of multi-task learning within the single-
modality domain of image, Ciresan et al. [58] applied the architec-
ture of deep CNNs to character recognition tasks for Latin and for
Chinese. The deep CNNs trained on Chinese characters are shown to
be easily capable of recognizing uppercase Latin letters. Further, learn-
ing Chinese characters is accelerated by first pre-training a CNN on a
small subset of all classes and then continuing to train on all classes.
 12
Conclusion

This monograph first presented a brief history of deep learning (focus-

ing on speech recognition) and developed a categorization scheme to
analyze the existing deep networks in the literature into unsupervised
(many of which are generative), supervised, and hybrid classes. The
deep autoencoder, the DSN (as well as many of its variants), and
the DBN–DNN or pre-trained DNN architectures, one in each of the
three classes, are discussed and analyzed in detail, as they appear to
be popular and promising approaches based on the authors’ personal
research experiences. Applications of deep learning in five broad
areas of information processing are also reviewed, including speech
and audio (Section 7), natural language modeling and processing
(Section 8), information retrieval (Section 9), object recognition and
computer vision (Section 10), and multi-modal and multi-task learning
(Section 11). There are other interesting yet non-mainstream applica-
tions of deep learning, which are not covered in this monograph. For
interested readers, please consult recent papers on the applications of
deep learning to optimal control in [219], to reinforcement learning in
[256], to malware classification in [66], to compressed sensing in [277],
to recognition confidence prediction in [173], to acoustic-articulatory
inversion mapping in [369], to emotion recognition from video in [189],

343
344 Conclusion

to emotion recognition from speech in [207, 222], to spoken language

understanding in [242, 366, 403], to speaker recognition in [351, 372],
to language-type recognition in [112], to dialogue state tracking for
spoken dialogue systems in [94, 152], to automatic voice activity
detection in [442], to speech enhancement in [396], to voice conversion
in [266], and to single-channel source separation in [132, 387].
The literature on deep learning is vast, mostly coming from
the machine learning community. The signal processing community
embraced deep learning only within the past four years or so (start-
ing around end of 2009) and the momentum is growing fast ever since.
This monograph is written mainly from the signal and information pro-
cessing perspective. Beyond surveying the existing deep learning work,
a classificatory scheme based on the architectures and on the nature
of the learning algorithms is developed, and an analysis and discus-
sions with concrete examples are presented. It is our hope that the
survey conducted in this monograph will provide insight for readers to
better understand the capability of the various deep learning systems
discussed in the monograph, the connection among different but sim-
ilar deep learning methods, and how to design proper deep learning
algorithms under different circumstances.
Throughout this review, the important message is conveyed that
building and learning deep hierarchies of features are highly desirable.
We have discussed the difficulty of learning parameters in all layers of
deep networks in one shot due to optimization difficulties that need
to be better understood. The unsupervised pre-training method in the
hybrid architecture of the DBN–DNN, which we reviewed in detail in
Section 5, appears to have offered a useful, albeit empirical, solution
to poor local optima in optimization and to regularization for the deep
model containing massive parameters even though a solid theoretical
foundation is still lacking. The effectiveness of the pre-training method,
which was one factor that stimulated the interest in deep learning by
the signal processing community in 2009 via collaborations between
academic and industrial researchers, is most prominent when the super-
vised training data are limited.
Deep learning is an emerging technology. Despite the empirical
promising results reported so far, much more work needs to be carried
 345

out. Importantly, it has not been the experience of deep learning

researchers that a single deep learning technique can be successful for
all classification tasks. For example, while the popular learning strat-
egy of generative pre-training followed by discriminative fine-tuning
seems to work well empirically for many tasks, it failed to work for
some other tasks that have been explored (e.g., language identifica-
tion or speaker recognition; unpublished). For these tasks, the features
extracted at the generative pre-training phase seem to describe the
underlying speech variations well but do not contain sufficient infor-
mation to distinguish between different languages. A learning strategy
that can extract discriminative yet also invariant features is expected to
provide better solutions. This idea has also been called “disentangling”
and is developed further in [24]. Further, extracting discriminative fea-
tures may greatly reduce the model size needed in many of the current
deep learning systems. Domain knowledge such as what kind of invari-
ance is useful for a specific task in hand (e.g., vision, speech, or natural
language) and what kind of regularization in terms of parameter con-
straints is key to the success of applying deep learning methods. More-
over, new types of DNN architectures and learning beyond the several
popular ones discussed in this monograph are currently under active
development by the deep learning research community (e.g., [24, 89]),
holding the promise to improve the performance of deep learning mod-
els in more challenging applications in signal processing and in artificial
intelligence.
Recent published work showed that there is vast room to improve
the current optimization techniques for learning deep architectures
[69, 208, 238, 239, 311, 356, 393]. To what extent pre-training is essen-
tial to learning the full set of parameters in deep architectures is
currently under investigation, especially when very large amounts of
labeled training data are available, reducing or even obliterating the
need for model regularization. Some preliminary results have been dis-
cussed in this monograph and in [55, 161, 323, 429].
In recent years, machine learning is becoming increasingly depen-
dent on large-scale data sets. For instance, many of the recent successes
of deep learning as discussed in this monograph have relied on the access
346 Conclusion

to massive data sets and massive computing power. It would become

increasingly difficult to explore the new algorithmic space without the
access to large, real-world data sets and without the related engineer-
ing expertise. How well deep learning algorithms behave would depend
heavily on the amount of data and computing power available. As we
showed with speech recognition examples, a deep learning algorithm
that appears to be performing not so remarkably on small data sets
can begin to perform considerably better when these limitations are
removed, one of main reasons for the recent resurgence in neural net-
work research. As an example, the DBN pre-training that ignited a new
era of (deep) machine learning research appears unnecessary if enough
data and computing power are used.
As a consequence, effective and scalable parallel algorithms are
critical for training deep models with large data sets, as in many com-
mon information processing applications such as speech recognition
and machine translation. The popular mini-batch stochastic gradient
technique is known to be difficult to parallelize over computers.
The common practice nowadays is to use GPGPUs to speed up the
learning process, although recent advance in developing asynchronous
stochastic gradient descent learning has shown promises by using
large-scale CPU clusters [69, 209] and GPU clusters [59]. In this
interesting computing architecture, many different replicas of the DNN
compute gradients on different subsets of the training data in parallel.
These gradients are communicated to a central parameter server
that updates the shared weights. Even though each replica typically
computes gradients using parameter values not immediately updated,
stochastic gradient descent is robust to the slight errors this has
introduced. To make deep learning techniques scalable to very large
training data, theoretically sound parallel learning and optimization
algorithms together with novel architectures need to be further devel-
oped [31, 39, 49, 69, 181, 322, 356]. Optimization methods specific to
speech recognition problems may need to be taken into account in order
to push speech recognition advances to the next level [46, 149, 393].
One major barrier to the application of DNNs and related deep
models is that it currently requires considerable skill and experience to
 347

choose sensible values for hyper-parameters such as the learning rate

schedule, the strength of the regularizer, the number of layers and the
number of units per layer, etc. Sensible values for one hyper-parameter
may depend on the values chosen for other hyper-parameters and
hyper-parameter tuning in DNNs is especially expensive. Some inter-
esting methods for solving the problem have been developed recently,
including random sampling [32] and Bayesian optimization procedure
[337]. Further research is needed in this important area.
This monograph, mainly in Sections 8 and 11 on natural language
and multi-modal applications, has touched on some recent work on
using deep learning methods to do reasoning, moving beyond the topic
of more straightforward pattern recognition using supervised, unsuper-
vised or hybrid learning methods to which much of this monograph
has been devoted to. In principle, since deep networks are naturally
equipped with distributed representations (rf. Table 3.1) using their
layer-wise collections of units for coding relations and coding entities,
concepts, events, topics, etc., they can potentially perform powerful
reasoning over structures, as argued in various historical publications
as well as recent essays [38, 156, 286, 288, 292, 336, 335]. While initial
explorations on this capability of deep networks have recently appeared
in the literature, as reviewed in Sections 8 and 11, much research is
needed. If successful, this new type of deep learning “machine” will
open up many novel and exciting applications in applied artificial intel-
ligence as a “thinking brain.” We expect growing work of deep learning
in this area, full of new challenges, in the future.
Further, solid theoretical foundations of deep learning need to be
established in a myriad of aspects. As an example, the success of deep
learning in unsupervised learning has not been demonstrated as much
as for supervised learning; yet the essence and major motivation of deep
learning lie right in unsupervised learning for automatically discover-
ing data representation. The issues involve appropriate objectives for
learning effective feature representations and the right deep learning
architectures/algorithms for distributed representations to effectively
disentangle the hidden explanatory factors of variation in the data.
Unfortunately, a majority of the successful deep learning techniques
348 Conclusion

have so far dealt with unstructured or “flat” classification problems.

For example, although speech recognition is a sequential classification
problem by nature, in the most successful and large-scale systems, a
separate HMM is used to handle the sequence structure and the DNN
is only used to produce the frame-level, unstructured posterior dis-
tributions. Recent proposals have called for and investigated moving
beyond the “flat” representations and incorporating structures in both
the deep learning architectures and input and output representations
[79, 136, 338, 349].
Finally, deep learning researchers have been advised by neuroscien-
tists to seriously consider a broader set of issues and learning architec-
tures so as to gain insight into biologically plausible representations in
the brain that may be useful for practical applications [272]. How can
computational neuroscience models about hierarchical brain structure
help improve engineering deep learning architectures? How may the
biologically feasible learning styles in the brain [158, 395] help design
more effective and more robust deep learning algorithms? All these
issues and those discussed earlier in this section will need intensive
research in order to further push the frontier of deep learning.
 References

[1] O. Abdel-Hamid, L. Deng, and D. Yu. Exploring convolutional neural

network structures and optimization for speech recognition. Proceedings
of Interspeech, 2013.
[2] O. Abdel-Hamid, L. Deng, D. Yu, and H. Jiang. Deep segmental neural
networks for speech recognition. In Proceedings of Interspeech. 2013.
[3] O. Abdel-Hamid, A. Mohamed, H. Jiang, and G. Penn. Applying convo-
lutional neural networks concepts to hybrid NN-HMM model for speech
recognition. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2012.
[4] A. Acero, L. Deng, T. Kristjansson, and J. Zhang. HMM adaptation
using vector taylor series for noisy speech recognition. In Proceedings
of Interspeech. 2000.
[5] G. Alain and Y. Bengio. What regularized autoencoders learn from the
data generating distribution. In Proceedings of International Conference
on Learning Representations (ICLR). 2013.
[6] G. Anthes. Deep learning comes of age. Communications of the Asso-
ciation for Computing Machinery (ACM), 56(6):13–15, June 2013.
[7] I. Arel, C. Rose, and T. Karnowski. Deep machine learning — a new
frontier in artificial intelligence. IEEE Computational Intelligence Mag-
azine, 5:13–18, November 2010.

349
350 References

[8] E. Arisoy, T. Sainath, B. Kingsbury, and B. Ramabhadran. Deep neural

network language models. In Proceedings of the Joint Human Language
Technology Conference and the North American Chapter of the Associ-
ation of Computational Linguistics (HLT-NAACL) Workshop. 2012.
[9] O. Aslan, H. Cheng, D. Schuurmans, and X. Zhang. Convex two-layer
modeling. In Proceedings of Neural Information Processing Systems
(NIPS). 2013.
[10] J. Ba and B. Frey. Adaptive dropout for training deep neural networks.
In Proceedings of Neural Information Processing Systems (NIPS). 2013.
[11] J. Baker, L. Deng, J. Glass, S. Khudanpur, C.-H. Lee, N. Morgan, and
D. O’Shaughnessy. Research developments and directions in speech
recognition and understanding. IEEE Signal Processing Magazine,
26(3):75–80, May 2009.
[12] J. Baker, L. Deng, J. Glass, S. Khudanpur, C.-H. Lee, N. Morgan, and
D. O’Shaughnessy. Updated MINS report on speech recognition and
understanding. IEEE Signal Processing Magazine, 26(4), July 2009.
[13] P. Baldi and P. Sadowski. Understanding dropout. In Proceedings of
Neural Information Processing Systems (NIPS). 2013.
[14] E. Battenberg, E. Schmidt, and J. Bello. Deep learning for
music, special session at International Conference on Acoustics
Speech and Signal Processing (ICASSP) (http://www.icassp2014.org/
special_sections.html#ss8), 2014.
[15] E. Batternberg and D. Wessel. Analyzing drum patterns using condi-
tional deep belief networks. In Proceedings of International Symposium
on Music Information Retrieval (ISMIR). 2012.
[16] P. Bell, P. Swietojanski, and S. Renals. Multi-level adaptive networks
in tandem and hybrid ASR systems. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2013.
[17] Y. Bengio. Artificial neural networks and their application to sequence
recognition. Ph.D. Thesis, McGill University, Montreal, Canada, 1991.
[18] Y. Bengio. New distributed probabilistic language models. Technical
Report, University of Montreal, 2002.
[19] Y. Bengio. Neural net language models. Scholarpedia, 3, 2008.
[20] Y. Bengio. Learning deep architectures for AI. in Foundations and
Trends in Machine Learning, 2(1):1–127, 2009.
References 351

[21] Y. Bengio. Deep learning of representations for unsupervised and trans-

fer learning. Journal of Machine Learning Research Workshop and Con-
ference Proceedings, 27:17–37, 2012.
[22] Y. Bengio. Deep learning of representations: Looking forward. In Sta-
tistical Language and Speech Processing, pages 1–37. Springer, 2013.
[23] Y. Bengio, N. Boulanger, and R. Pascanu. Advances in optimizing recur-
rent networks. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013.
[24] Y. Bengio, A. Courville, and P. Vincent. Representation learning: A
review and new perspectives. IEEE Transactions on Pattern Analysis
and Machine Intelligence (PAMI), 38:1798–1828, 2013.
[25] Y. Bengio, R. De Mori, G. Flammia, and R. Kompe. Global optimiza-
tion of a neural network-hidden markov model hybrid. IEEE Transac-
tions on Neural Networks, 3:252–259, 1992.
[26] Y. Bengio, R. Ducharme, P. Vincent, and C. Jauvin. A neural proba-
bilistic language model. In Proceedings of Neural Information Processing
Systems (NIPS). 2000.
[27] Y. Bengio, R. Ducharme, P. Vincent, and C. Jauvin. A neural proba-
bilistic language model. Journal of Machine Learning Research, 3:1137–
1155, 2003.
[28] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-
wise training of deep networks. In Proceedings of Neural Information
Processing Systems (NIPS). 2006.
[29] Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependen-
cies with gradient descent is difficult. IEEE Transactions on Neural
Networks, 5:157–166, 1994.
[30] Y. Bengio, E. Thibodeau-Laufer, and J. Yosinski. Deep generative
stochastic networks trainable by backprop. arXiv 1306:1091, 2013.
also accepted to appear in Proceedings of International Conference on
Machine Learning (ICML), 2014.
[31] Y. Bengio, L. Yao, G. Alain, and P. Vincent. Generalized denoising
autoencoders as generative models. In Proceedings of Neural Informa-
tion Processing Systems (NIPS). 2013.
[32] J. Bergstra and Y. Bengio. Random search for hyper-parameter opti-
mization. Journal on Machine Learning Research, 3:281–305, 2012.
352 References

[33] A. Biem, S. Katagiri, E. McDermott, and B. Juang. An application

of discriminative feature extraction to filter-bank-based speech recog-
nition. IEEE Transactions on Speech and Audio Processing, 9:96–110,
2001.
[34] J. Bilmes. Dynamic graphical models. IEEE Signal Processing Maga-
zine, 33:29–42, 2010.
[35] J. Bilmes and C. Bartels. Graphical model architectures for speech
recognition. IEEE Signal Processing Magazine, 22:89–100, 2005.
[36] A. Bordes, X. Glorot, J. Weston, and Y. Bengio. A semantic matching
energy function for learning with multi-relational data — application
to word-sense disambiguation. Machine Learning, May 2013.
[37] A. Bordes, J. Weston, R. Collobert, and Y. Bengio. Learning structured
embeddings of knowledge bases. In Proceedings of Association for the
Advancement of Artificial Intelligence (AAAI). 2011.
[38] L. Bottou. From machine learning to machine reasoning: An essay.
Journal of Machine Learning Research, 14:3207–3260, 2013.
[39] L. Bottou and Y. LeCun. Large scale online learning. In Proceedings of
Neural Information Processing Systems (NIPS). 2004.
[40] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent. Modeling
Temporal dependencies in high-dimensional sequences: Application to
polyphonic music generation and transcription. In Proceedings of Inter-
national Conference on Machine Learning (ICML). 2012.
[41] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent. Audio chord
recognition with recurrent neural networks. In Proceedings of Interna-
tional Symposium on Music Information Retrieval (ISMIR). 2013.
[42] H. Bourlard and N. Morgan. Connectionist Speech Recognition: A
Hybrid Approach. Kluwer, Norwell, MA, 1993.
[43] J. Bouvrie. Hierarchical learning: Theory with applications in speech
and vision. Ph.D. thesis, MIT, 2009.
[44] L. Breiman. Stacked regression. Machine Learning, 24:49–64, 1996.
[45] J. Bridle, L. Deng, J. Picone, H. Richards, J. Ma, T. Kamm, M. Schus-
ter, S. Pike, and R. Reagan. An investigation of segmental hidden
dynamic models of speech coarticulation for automatic speech recogni-
tion. Final Report for 1998 Workshop on Language Engineering, CLSP,
Johns Hopkins, 1998.
[46] P. Cardinal, P. Dumouchel, and G. Boulianne. Large vocabulary speech
recognition on parallel architectures. IEEE Transactions on Audio,
Speech, and Language Processing, 21(11):2290–2300, November 2013.
References 353

[47] R. Caruana. Multitask learning. Machine Learning, 28:41–75, 1997.

[48] J. Chen and L. Deng. A primal-dual method for training recurrent
neural networks constrained by the echo-state property. In Proceedings
of International Conference on Learning Representations. April 2014.
[49] X. Chen, A. Eversole, G. Li, D. Yu, and F. Seide. Pipelined back-
propagation for context-dependent deep neural networks. In Proceedings
of Interspeech. 2012.
[50] R. Chengalvarayan and L. Deng. Hmm-based speech recognition using
state-dependent, discriminatively derived transforms on Mel-warped
DFT features. IEEE Transactions on Speech and Audio Processing,
pages 243–256, 1997.
[51] R. Chengalvarayan and L. Deng. Use of generalized dynamic feature
parameters for speech recognition. IEEE Transactions on Speech and
Audio Processing, pages 232–242, 1997a.
[52] R. Chengalvarayan and L. Deng. Speech trajectory discrimination using
the minimum classification error learning. IEEE Transactions on Speech
and Audio Processing, 6(6):505–515, 1998.
[53] Y. Cho and L. Saul. Kernel methods for deep learning. In Proceedings of
Neural Information Processing Systems (NIPS), pages 342–350. 2009.
[54] D. Ciresan, A. Giusti, L. Gambardella, and J. Schmidhuber. Deep neural
networks segment neuronal membranes in electron microscopy images.
In Proceedings of Neural Information Processing Systems (NIPS). 2012.
[55] D. Ciresan, U. Meier, L. Gambardella, and J. Schmidhuber. Deep, big,
simple neural nets for handwritten digit recognition. Neural Computa-
tion, December 2010.
[56] D. Ciresan, U. Meier, J. Masci, and J. Schmidhuber. A committee of
neural networks for traffic sign classification. In Proceedings of Interna-
tional Joint Conference on Neural Networks (IJCNN). 2011.
[57] D. Ciresan, U. Meier, and J. Schmidhuber. Multi-column deep neural
networks for image classification. In Proceedings of Computer Vision
and Pattern Recognition (CVPR). 2012.
[58] D. C. Ciresan, U. Meier, and J. Schmidhuber. Transfer learning for Latin
and Chinese characters with deep neural networks. In Proceedings of
International Joint Conference on Neural Networks (IJCNN). 2012.
[59] A. Coates, B. Huval, T. Wang, D. Wu, A. Ng, and B. Catanzaro. Deep
learning with COTS HPC. In Proceedings of International Conference
on Machine Learning (ICML). 2013.
354 References

[60] W. Cohen and R. V. de Carvalho. Stacked sequential learning. In

Proceedings of International Joint Conference on Artificial Intelligence
(IJCAI), pages 671–676. 2005.
[61] R. Collobert. Deep learning for efficient discriminative parsing. In
Proceedings of Artificial Intelligence and Statistics (AISTATS). 2011.
[62] R. Collobert and J. Weston. A unified architecture for natural language
processing: Deep neural networks with multitask learning. In Proceed-
ings of International Conference on Machine Learning (ICML). 2008.
[63] R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, and
P. Kuksa. Natural language processing (almost) from scratch. Journal
on Machine Learning Research, 12:2493–2537, 2011.
[64] G. Dahl, M. Ranzato, A. Mohamed, and G. Hinton. Phone recognition
with the mean-covariance restricted boltzmann machine. In Proceedings
of Neural Information Processing Systems (NIPS), volume 23, pages
469–477. 2010.
[65] G. Dahl, T. Sainath, and G. Hinton. Improving deep neural networks
for LVCSR using rectified linear units and dropout. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2013.
[66] G. Dahl, J. Stokes, L. Deng, and D. Yu. Large-scale malware classifi-
cation using random projections and neural networks. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2013.
[67] G. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent DBN-
HMMs in large vocabulary continuous speech recognition. In Proceed-
ings of International Conference on Acoustics Speech and Signal Pro-
cessing (ICASSP). 2011.
[68] G. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent, pre-trained
deep neural networks for large vocabulary speech recognition. IEEE
Transactions on Audio, Speech, & Language Processing, 20(1):30–42,
January 2012.
[69] J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, Q. Le, M. Mao,
M. Ranzato, A. Senior, P. Tucker, K. Yang, and A. Ng. Large scale
distributed deep networks. In Proceedings of Neural Information Pro-
cessing Systems (NIPS). 2012.
[70] K. Demuynck and F. Triefenbach. Porting concepts from DNNs back
to GMMs. In Proceedings of the Automatic Speech Recognition and
Understanding Workshop (ASRU). 2013.
References 355

[71] L. Deng. A generalized hidden Markov model with state-conditioned

trend functions of time for the speech signal. Signal Processing,
27(1):65–78, 1992.
[72] L. Deng. A stochastic model of speech incorporating hierarchical nonsta-
tionarity. IEEE Transactions on Speech and Audio Processing, 1(4):471–
475, 1993.
[73] L. Deng. A dynamic, feature-based approach to the interface between
phonology and phonetics for speech modeling and recognition. Speech
Communication, 24(4):299–323, 1998.
[74] L. Deng. Computational models for speech production. In Compu-
tational Models of Speech Pattern Processing, pages 199–213. Springer
Verlag, 1999.
[75] L. Deng. Switching dynamic system models for speech articulation and
acoustics. In Mathematical Foundations of Speech and Language Pro-
cessing, pages 115–134. Springer-Verlag, New York, 2003.
[76] L. Deng. Dynamic Speech Models — Theory, Algorithm, and Applica-
tion. Morgan & Claypool, December 2006.
[77] L. Deng. An overview of deep-structured learning for information pro-
cessing. In Proceedings of Asian-Pacific Signal & Information Process-
ing Annual Summit and Conference (APSIPA-ASC). October 2011.
[78] L. Deng. The MNIST database of handwritten digit images for machine
learning research. IEEE Signal Processing Magazine, 29(6), November
2012.
[79] L. Deng. Design and learning of output representations for speech recog-
nition. In Neural Information Processing Systems (NIPS) Workshop on
Learning Output Representations. December 2013.
[80] L. Deng. A tutorial survey of architectures, algorithms, and applications
for deep learning. In Asian-Pacific Signal & Information Processing
Association Transactions on Signal and Information Processing. 2013.
[81] L. Deng, O. Abdel-Hamid, and D. Yu. A deep convolutional neural
network using heterogeneous pooling for trading acoustic invariance
with phonetic confusion. In Proceedings of International Conference
on Acoustics Speech and Signal Processing (ICASSP). 2013.
[82] L. Deng, A. Acero, L. Jiang, J. Droppo, and X. Huang. High perfor-
mance robust speech recognition using stereo training data. In Pro-
ceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2001.
356 References

[83] L. Deng and M. Aksmanovic. Speaker-independent phonetic classifi-

cation using hidden markov models with state-conditioned mixtures of
trend functions. IEEE Transactions on Speech and Audio Processing,
5:319–324, 1997.
[84] L. Deng, M. Aksmanovic, D. Sun, and J. Wu. Speech recognition using
hidden Markov models with polynomial regression functions as nonsta-
tionary states. IEEE Transactions on Speech and Audio Processing,
2(4):507–520, 1994.
[85] L. Deng and J. Chen. Sequence classification using the high-level fea-
tures extracted from deep neural networks. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2014.
[86] L. Deng and K. Erler. Structural design of a hidden Markov model
based speech recognizer using multi-valued phonetic features: Compar-
ison with segmental speech units. Journal of the Acoustical Society of
America, 92(6):3058–3067, 1992.
[87] L. Deng, K. Hassanein, and M. Elmasry. Analysis of correlation struc-
ture for a neural predictive model with application to speech recognition.
Neural Networks, 7(2):331–339, 1994.
[88] L. Deng, X. He, and J. Gao. Deep stacking networks for informa-
tion retrieval. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013c.
[89] L. Deng, G. Hinton, and B. Kingsbury. New types of deep neural
network learning for speech recognition and related applications: An
overview. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013b.
[90] L. Deng and X. D. Huang. Challenges in adopting speech recognition.
Communications of the Association for Computing Machinery (ACM),
47(1):11–13, January 2004.
[91] L. Deng, B. Hutchinson, and D. Yu. Parallel training of deep stacking
networks. In Proceedings of Interspeech. 2012b.
[92] L. Deng, M. Lennig, V. Gupta, F. Seitz, P. Mermelstein, and P. Kenny.
Phonemic hidden Markov models with continuous mixture output den-
sities for large vocabulary word recognition. IEEE Transactions on
Signal Processing, 39(7):1677–1681, 1991.
[93] L. Deng, M. Lennig, F. Seitz, and P. Mermelstein. Large vocabulary
word recognition using context-dependent allophonic hidden Markov
models. Computer Speech and Language, 4(4):345–357, 1990.
References 357

[94] L. Deng, J. Li, K. Huang, Yao, D. Yu, F. Seide, M. Seltzer, G. Zweig,

X. He, J. Williams, Y. Gong, and A. Acero. Recent advances in deep
learning for speech research at Microsoft. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2013a.
[95] L. Deng and X. Li. Machine learning paradigms in speech recogni-
tion: An overview. IEEE Transactions on Audio, Speech, & Language,
21:1060–1089, May 2013.
[96] L. Deng and J. Ma. Spontaneous speech recognition using a statistical
coarticulatory model for the vocal tract resonance dynamics. Journal
of the Acoustical Society America, 108:3036–3048, 2000.
[97] L. Deng and D. O’Shaughnessy. Speech Processing — A Dynamic and
Optimization-Oriented Approach. Marcel Dekker, 2003.
[98] L. Deng, G. Ramsay, and D. Sun. Production models as a structural
basis for automatic speech recognition. Speech Communication, 33(2–
3):93–111, August 1997.
[99] L. Deng and H. Sameti. Transitional speech units and their represen-
tation by regressive Markov states: Applications to speech recognition.
IEEE Transactions on speech and audio processing, 4(4):301–306, July
1996.
[100] L. Deng, M. Seltzer, D. Yu, A. Acero, A. Mohamed, and G. Hinton.
Binary coding of speech spectrograms using a deep autoencoder. In
Proceedings of Interspeech. 2010.
[101] L. Deng and D. Sun. A statistical approach to automatic speech
recognition using the atomic speech units constructed from overlap-
ping articulatory features. Journal of the Acoustical Society of America,
85(5):2702–2719, 1994.
[102] L. Deng, G. Tur, X. He, and D. Hakkani-Tur. Use of kernel deep convex
networks and end-to-end learning for spoken language understanding.
In Proceedings of IEEE Workshop on Spoken Language Technologies.
December 2012.
[103] L. Deng, K. Wang, A. Acero, H. W. Hon, J. Droppo, C. Boulis, Y. Wang,
D. Jacoby, M. Mahajan, C. Chelba, and X. Huang. Distributed speech
processing in mipad’s multimodal user interface. IEEE Transactions on
Speech and Audio Processing, 10(8):605–619, 2002.
[104] L. Deng, J. Wu, J. Droppo, and A. Acero. Dynamic compensation of
HMM variances using the feature enhancement uncertainty computed
from a parametric model of speech distortion. IEEE Transactions on
Speech and Audio Processing, 13(3):412–421, 2005.
358 References

[105] L. Deng and D. Yu. Use of differential cepstra as acoustic features

in hidden trajectory modeling for phonetic recognition. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2007.
[106] L. Deng and D. Yu. Deep convex network: A scalable architecture for
speech pattern classification. In Proceedings of Interspeech. 2011.
[107] L. Deng, D. Yu, and A. Acero. A bidirectional target filtering model of
speech coarticulation: Two-stage implementation for phonetic recogni-
tion. IEEE Transactions on Audio and Speech Processing, 14(1):256–
265, January 2006.
[108] L. Deng, D. Yu, and A. Acero. Structured speech modeling. IEEE
Transactions on Audio, Speech and Language Processing, 14(5):1492–
1504, September 2006.
[109] L. Deng, D. Yu, and G. Hinton. Deep learning for speech recognition and
related applications. Neural Information Processing Systems (NIPS)
Workshop, 2009.
[110] L. Deng, D. Yu, and J. Platt. Scalable stacking and learning for build-
ing deep architectures. In Proceedings of International Conference on
Acoustics Speech and Signal Processing (ICASSP). 2012a.
[111] T. Deselaers, S. Hasan, O. Bender, and H. Ney. A deep learning
approach to machine transliteration. In Proceedings of 4th Workshop on
Statistical Machine Translation, pages 233–241. Athens, Greece, March
2009.
[112] A. Diez. Automatic language recognition using deep neural networks.
Thesis, Universidad Autonoma de Madrid, SPAIN, September 2013.
[113] P. Dognin and V. Goel. Combining stochastic average gradient and
hessian-free optimization for sequence training of deep neural networks.
In Proceedings of the Automatic Speech Recognition and Understanding
Workshop (ASRU). 2013.
[114] D. Erhan, Y. Bengio, A. Courvelle, P. Manzagol, P. Vencent, and S. Ben-
gio. Why does unsupervised pre-training help deep learning? Journal
on Machine Learning Research, pages 201–208, 2010.
[115] R. Fernandez, A. Rendel, B. Ramabhadran, and R. Hoory. F0 contour
prediction with a deep belief network-gaussian process hybrid model. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP), pages 6885–6889. 2013.
[116] S. Fine, Y. Singer, and N. Tishby. The hierarchical hidden Markov
model: Analysis and applications. Machine Learning, 32:41–62, 1998.
References 359

[117] A. Frome, G. Corrado, J. Shlens, S. Bengio, J. Dean, M. Ranzato, and

T. Mikolov. Devise: A deep visual-semantic embedding model. In Pro-
ceedings of Neural Information Processing Systems (NIPS). 2013.
[118] Q. Fu, X. He, and L. Deng. Phone-discriminating minimum classifica-
tion error (p-mce) training for phonetic recognition. In Proceedings of
Interspeech. 2007.
[119] M. Gales. Model-based approaches to handling uncertainty. In Robust
Speech Recognition of Uncertain or Missing Data: Theory and Applica-
tion, pages 101–125. Springer, 2011.
[120] J. Gao, X. He, and J.-Y. Nie. Clickthrough-based translation models
for web search: From word models to phrase models. In Proceedings of
Conference on Information and Knowledge Management (CIKM). 2010.
[121] J. Gao, X. He, W. Yih, and L. Deng. Learning semantic representations
for the phrase translation model. In Proceedings of Neural Informa-
tion Processing Systems (NIPS) Workshop on Deep Learning. December
2013.
[122] J. Gao, X. He, W. Yih, and L. Deng. Learning semantic representations
for the phrase translation model. MSR-TR-2013-88, September 2013.
[123] J. Gao, X. He, W. Yih, and L. Deng. Learning continuous phrase rep-
resentations for translation modeling. In Proceedings of Association for
Computational Linguistics (ACL). 2014.
[124] J. Gao, K. Toutanova, and W.-T. Yih. Clickthrough-based latent seman-
tic models for web search. In Proceedings of Special Interest Group on
Information Retrieval (SIGIR). 2011.
[125] R. Gens and P. Domingo. Discriminative learning of sum-product net-
works. Neural Information Processing Systems (NIPS), 2012.
[126] D. George. How the brain might work: A hierarchical and temporal
model for learning and recognition. Ph.D. thesis, Stanford University,
2008.
[127] M. Gibson and T. Hain. Error approximation and minimum phone error
acoustic model estimation. IEEE Transactions on Audio, Speech, and
Language Processing, 18(6):1269–1279, August 2010.
[128] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature
hierarchies for accurate object detection and semantic segmentation.
arXiv:1311.2524v1, 2013.
[129] X. Glorot and Y. Bengio. Understanding the difficulty of training deep
feed-forward neural networks. In Proceedings of Artificial Intelligence
and Statistics (AISTATS). 2010.
360 References

[130] X. Glorot, A. Bordes, and Y. Bengio. Deep sparse rectifier neural

networks. In Proceedings of Artificial Intelligence and Statistics (AIS-
TATS). April 2011.
[131] I. Goodfellow, M. Mirza, A. Courville, and Y. Bengio. Multi-prediction
deep boltzmann machines. In Proceedings of Neural Information Pro-
cessing Systems (NIPS). 2013.
[132] E. Grais, M. Sen, and H. Erdogan. Deep neural networks for single
channel source separation. arXiv:1311.2746v1, 2013.
[133] A. Graves. Sequence transduction with recurrent neural networks. Rep-
resentation Learning Workshop, International Conference on Machine
Learning (ICML), 2012.
[134] A. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber. Connection-
ist temporal classification: Labeling unsegmented sequence data with
recurrent neural networks. In Proceedings of International Conference
on Machine Learning (ICML). 2006.
[135] A. Graves, N. Jaitly, and A. Mohamed. Hybrid speech recognition with
deep bidirectional LSTM. In Proceedings of the Automatic Speech Recog-
nition and Understanding Workshop (ASRU). 2013.
[136] A. Graves, A. Mohamed, and G. Hinton. Speech recognition with deep
recurrent neural networks. In Proceedings of International Conference
on Acoustics Speech and Signal Processing (ICASSP). 2013.
[137] F. Grezl and P. Fousek. Optimizing bottle-neck features for LVCSR. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2008.
[138] C. Gulcehre, K. Cho, R. Pascanu, and Y. Bengio. Learned-
norm pooling for deep feedforward and recurrent neural networks.
http://arxiv.org/abs/1311.1780, 2014.
[139] M. Gutmann and A. Hyvarinen. Noise-contrastive estimation of unnor-
malized statistical models, with applications to natural image statistics.
Journal of Machine Learning Research, 13:307–361, 2012.
[140] T. Hain, L. Burget, J. Dines, P. Garner, F. Grezl, A. Hannani, M. Hui-
jbregts, M. Karafiat, M. Lincoln, and V. Wan. Transcribing meetings
with the AMIDA systems. IEEE Transactions on Audio, Speech, and
Language Processing, 20:486–498, 2012.
[141] P. Hamel and D. Eck. Learning features from music audio with deep
belief networks. In Proceedings of International Symposium on Music
Information Retrieval (ISMIR). 2010.
References 361

[142] G. Hawkins, S. Ahmad, and D. Dubinsky. Hierarchical temporal mem-

ory including HTM cortical learning algorithms. Numenta Technical
Report, December 10 2010.
[143] J. Hawkins and S. Blakeslee. On Intelligence: How a New Understanding
of the Brain will lead to the Creation of Truly Intelligent Machines.
Times Books, New York, 2004.
[144] X. He and L. Deng. Speech recognition, machine translation, and speech
translation — a unifying discriminative framework. IEEE Signal Pro-
cessing Magazine, 28, November 2011.
[145] X. He and L. Deng. Optimization in speech-centric information process-
ing: Criteria and techniques. In Proceedings of International Conference
on Acoustics Speech and Signal Processing (ICASSP). 2012.
[146] X. He and L. Deng. Speech-centric information processing: An
optimization-oriented approach. In Proceedings of the IEEE. 2013.
[147] X. He, L. Deng, and W. Chou. Discriminative learning in sequential pat-
tern recognition — a unifying review for optimization-oriented speech
recognition. IEEE Signal Processing Magazine, 25:14–36, 2008.
[148] G. Heigold, H. Ney, P. Lehnen, T. Gass, and R. Schluter. Equivalence of
generative and log-liner models. IEEE Transactions on Audio, Speech,
and Language Processing, 19(5):1138–1148, February 2011.
[149] G. Heigold, H. Ney, and R. Schluter. Investigations on an EM-style opti-
mization algorithm for discriminative training of HMMs. IEEE Trans-
actions on Audio, Speech, and Language Processing, 21(12):2616–2626,
December 2013.
[150] G. Heigold, V. Vanhoucke, A. Senior, P. Nguyen, M. Ranzato, M. Devin,
and J. Dean. Multilingual acoustic models using distributed deep neu-
ral networks. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013.
[151] I. Heintz, E. Fosler-Lussier, and C. Brew. Discriminative input stream
combination for conditional random field phone recognition. IEEE
Transactions on Audio, Speech, and Language Processing, 17(8):1533–
1546, November 2009.
[152] M. Henderson, B. Thomson, and S. Young. Deep neural network
approach for the dialog state tracking challenge. In Proceedings of Spe-
cial Interest Group on Disclosure and Dialogue (SIGDIAL). 2013.
[153] M. Hermans and B. Schrauwen. Training and analysing deep recur-
rent neural networks. In Proceedings of Neural Information Processing
Systems (NIPS). 2013.
362 References

[154] H. Hermansky, D. Ellis, and S. Sharma. Tandem connectionist feature

extraction for conventional HMM systems. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2000.
[155] Y. Hifny and S. Renals. Speech recognition using augmented conditional
random fields. IEEE Transactions on Audio, Speech, and Language
Processing, 17(2):354–365, February 2009.
[156] G. Hinton. Mapping part-whole hierarchies into connectionist networks.
Artificial Intelligence, 46:47–75, 1990.
[157] G. Hinton. Preface to the special issue on connectionist symbol pro-
cessing. Artificial Intelligence, 46:1–4, 1990.
[158] G. Hinton. The ups and downs of Hebb synapses. Canadian Psychology,
44:10–13, 2003.
[159] G. Hinton. A practical guide to training restricted boltzmann machines.
UTML Tech Report 2010-003, Univ. Toronto, August 2010.
[160] G. Hinton. A better way to learn features. Communications of the
Association for Computing Machinery (ACM), 54(10), October 2011.
[161] G. Hinton, L. Deng, D. Yu, G. Dahl, A. Mohamed, N. Jaitly, A. Senior,
V. Vanhoucke, P. Nguyen, T. Sainath, and B. Kingsbury. Deep neu-
ral networks for acoustic modeling in speech recognition. IEEE Signal
Processing Magazine, 29(6):82–97, November 2012.
[162] G. Hinton, A. Krizhevsky, and S. Wang. Transforming autoencoders. In
Proceedings of International Conference on Artificial Neural Networks.
2011.
[163] G. Hinton, S. Osindero, and Y. Teh. A fast learning algorithm for deep
belief nets. Neural Computation, 18:1527–1554, 2006.
[164] G. Hinton and R. Salakhutdinov. Reducing the dimensionality of data
with neural networks. Science, 313(5786):504–507, July 2006.
[165] G. Hinton and R. Salakhutdinov. Discovering binary codes for docu-
ments by learning deep generative models. Topics in Cognitive Science,
pages 1–18, 2010.
[166] G. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. Salakhut-
dinov. Improving neural networks by preventing co-adaptation of fea-
ture detectors. arXiv: 1207.0580v1, 2012.
[167] S. Hochreiter. Untersuchungen zu dynamischen neuronalen net-
zen. Diploma thesis, Institut fur Informatik, Technische Universitat
Munchen, 1991.
References 363

[168] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural

Computation, 9:1735–1780, 1997.
[169] E. Huang, R. Socher, C. Manning, and A. Ng. Improving word represen-
tations via global context and multiple word prototypes. In Proceedings
of Association for Computational Linguistics (ACL). 2012.
[170] J. Huang, J. Li, L. Deng, and D. Yu. Cross-language knowledge transfer
using multilingual deep neural networks with shared hidden layers. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2013.
[171] P. Huang, L. Deng, M. Hasegawa-Johnson, and X. He. Random fea-
tures for kernel deep convex network. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2013.
[172] P. Huang, X. He, J. Gao, L. Deng, A. Acero, and L. Heck. Learning
deep structured semantic models for web search using clickthrough data.
Association for Computing Machinery (ACM) International Conference
Information and Knowledge Management (CIKM), 2013.
[173] P. Huang, K. Kumar, C. Liu, Y. Gong, and L. Deng. Predicting speech
recognition confidence using deep learning with word identity and score
features. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2013.
[174] S. Huang and S. Renals. Hierarchical bayesian language models for
conversational speech recognition. IEEE Transactions on Audio, Speech,
and Language Processing, 18(8):1941–1954, November 2010.
[175] X. Huang, A. Acero, C. Chelba, L. Deng, J. Droppo, D. Duchene,
J. Goodman, and H. Hon. Mipad: A multimodal interaction proto-
type. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2001.
[176] Y. Huang, D. Yu, Y. Gong, and C. Liu. Semi-supervised GMM and DNN
acoustic model training with multi-system combination and confidence
re-calibration. In Proceedings of Interspeech, pages 2360–2364. 2013.
[177] E. Humphrey and J. Bello. Rethinking automatic chord recognition
with convolutional neural networks. In Proceedings of International
Conference on Machine Learning and Application (ICMLA). 2012a.
[178] E. Humphrey, J. Bello, and Y. LeCun. Moving beyond feature design:
Deep architectures and automatic feature learning in music informat-
ics. In Proceedings of International Symposium on Music Information
Retrieval (ISMIR). 2012.
364 References

[179] E. Humphrey, J. Bello, and Y. LeCun. Feature learning and deep archi-
tectures: New directions for music informatics. Journal of Intelligent
Information Systems, 2013.
[180] B. Hutchinson, L. Deng, and D. Yu. A deep architecture with bilinear
modeling of hidden representations: Applications to phonetic recogni-
tion. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2012.
[181] B. Hutchinson, L. Deng, and D. Yu. Tensor deep stacking net-
works. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, 35:1944–1957, 2013.
[182] D. Imseng, P. Motlicek, P. Garner, and H. Bourlard. Impact of deep
MLP architecture on different modeling techniques for under-resourced
speech recognition. In Proceedings of the Automatic Speech Recognition
and Understanding Workshop (ASRU). 2013.
[183] N. Jaitly and G. Hinton. Learning a better representation of speech
sound waves using restricted boltzmann machines. In Proceedings of
International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2011.
[184] N. Jaitly, P. Nguyen, and V. Vanhoucke. Application of pre-trained deep
neural networks to large vocabulary speech recognition. In Proceedings
of Interspeech. 2012.
[185] K. Jarrett, K. Kavukcuoglu, and Y. LeCun. What is the best multi-
stage architecture for object recognition? In Proceedings of International
Conference on Computer Vision, pages 2146–2153. 2009.
[186] H. Jiang and X. Li. Parameter estimation of statistical models using
convex optimization: An advanced method of discriminative training
for speech and language processing. IEEE Signal Processing Magazine,
27(3):115–127, 2010.
[187] B. Juang, S. Levinson, and M. Sondhi. Maximum likelihood estimation
for multivariate mixture observations of Markov chains. IEEE Trans-
actions on Information Theory, 32:307–309, 1986.
[188] B.-H. Juang, W. Chou, and C.-H. Lee. Minimum classification error
rate methods for speech recognition. IEEE Transactions On Speech
and Audio Processing, 5:257–265, 1997.
[189] S. Kahou et al. Combining modality specific deep neural networks for
emotion recognition in video. In Proceedings of International Conference
on Multimodal Interaction (ICMI). 2013.
References 365

[190] S. Kang, X. Qian, and H. Meng. Multi-distribution deep belief network

for speech synthesis. In Proceedings of International Conference on
Acoustics Speech and Signal Processing (ICASSP), pages 8012–8016.
2013.
[191] Y. Kashiwagi, D. Saito, N. Minematsu, and K. Hirose. Discriminative
piecewise linear transformation based on deep learning for noise robust
automatic speech recognition. In Proceedings of the Automatic Speech
Recognition and Understanding Workshop (ASRU). 2013.
[192] K. Kavukcuoglu, P. Sermanet, Y. Boureau, K. Gregor, M. Mathieu,
and Y. LeCun. Learning convolutional feature hierarchies for visual
recognition. In Proceedings of Neural Information Processing Systems
(NIPS). 2010.
[193] H. Ketabdar and H. Bourlard. Enhanced phone posteriors for improving
speech recognition systems. IEEE Transactions on Audio, Speech, and
Language Processing, 18(6):1094–1106, August 2010.
[194] B. Kingsbury. Lattice-based optimization of sequence classification cri-
teria for neural-network acoustic modeling. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2009.
[195] B. Kingsbury, T. Sainath, and H. Soltau. Scalable minimum bayes
risk training of deep neural network acoustic models using distributed
hessian-free optimization. In Proceedings of Interspeech. 2012.
[196] R. Kiros, R. Zemel, and R. Salakhutdinov. Multimodal neural lan-
guage models. In Proceedings of Neural Information Processing Systems
(NIPS) Deep Learning Workshop. 2013.
[197] T. Ko and B. Mak. Eigentriphones for context-dependent acoustic mod-
eling. IEEE Transactions on Audio, Speech, and Language Processing,
21(6):1285–1294, 2013.
[198] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with
deep convolutional neural networks. In Proceedings of Neural Informa-
tion Processing Systems (NIPS). 2012.
[199] Y. Kubo, T. Hori, and A. Nakamura. Integrating deep neural networks
into structural classification approach based on weighted finite-state
transducers. In Proceedings of Interspeech. 2012.
[200] R. Kurzweil. How to Create a Mind. Viking Books, December 2012.
[201] P. Lal and S. King. Cross-lingual automatic speech recognition using
tandem features. IEEE Transactions on Audio, Speech, and Language
Processing, 21(12):2506–2515, December 2013.
366 References

[202] K. Lang, A. Waibel, and G. Hinton. A time-delay neural network archi-

tecture for isolated word recognition. Neural Networks, 3(1):23–43, 1990.
[203] H. Larochelle and Y. Bengio. Classification using discriminative
restricted boltzmann machines. In Proceedings of International Con-
ference on Machine Learning (ICML). 2008.
[204] D. Le and P. Mower. Emotion recognition from spontaneous speech
using hidden markov models with deep belief networks. In Proceed-
ings of the Automatic Speech Recognition and Understanding Workshop
(ASRU). 2013.
[205] H. Le, A. Allauzen, G. Wisniewski, and F. Yvon. Training continuous
space language models: Some practical issues. In Proceedings of Empiri-
cal Methods in Natural Language Processing (EMNLP), pages 778–788.
2010.
[206] H. Le, I. Oparin, A. Allauzen, J. Gauvain, and F. Yvon. Structured
output layer neural network language model. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2011.
[207] H. Le, I. Oparin, A. Allauzen, J.-L. Gauvain, and F. Yvon. Struc-
tured output layer neural network language models for speech recogni-
tion. IEEE Transactions on Audio, Speech, and Language Processing,
21(1):197–206, January 2013.
[208] Q. Le, J. Ngiam, A. Coates, A. Lahiri, B. Prochnow, and A. Ng. On
optimization methods for deep learning. In Proceedings of International
Conference on Machine Learning (ICML). 2011.
[209] Q. Le, M. Ranzato, R. Monga, M. Devin, G. Corrado, K. Chen, J. Dean,
and A. Ng. Building high-level features using large scale unsupervised
learning. In Proceedings of International Conference on Machine Learn-
ing (ICML). 2012.
[210] Y. LeCun. Learning invariant feature hierarchies. In Proceedings of
European Conference on Computer Vision (ECCV). 2012.
[211] Y. LeCun and Y. Bengio. Convolutional networks for images, speech,
and time series. In M. Arbib, editor, The Handbook of Brain The-
ory and Neural Networks, pages 255–258. MIT Press, Cambridge, Mas-
sachusetts, 1995.
[212] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learn-
ing applied to document recognition. Proceedings of the IEEE, 86:2278–
2324, 1998.
References 367

[213] Y. LeCun, S. Chopra, M. Ranzato, and F. Huang. Energy-based models

in document recognition and computer vision. In Proceedings of Inter-
national Conference on Document Analysis and Recognition (ICDAR).
2007.
[214] C.-H. Lee. From knowledge-ignorant to knowledge-rich modeling: A new
speech research paradigm for next-generation automatic speech recog-
nition. In Proceedings of International Conference on Spoken Language
Processing (ICSLP), pages 109–111. 2004.
[215] H. Lee, R. Grosse, R. Ranganath, and A. Ng. Convolutional deep belief
networks for scalable unsupervised learning of hierarchical representa-
tions. In Proceedings of International Conference on Machine Learning
(ICML). 2009.
[216] H. Lee, R. Grosse, R. Ranganath, and A. Ng. Unsupervised learning
of hierarchical representations with convolutional deep belief networks.
Communications of the Association for Computing Machinery (ACM),
54(10):95–103, October 2011.
[217] H. Lee, Y. Largman, P. Pham, and A. Ng. Unsupervised feature learning
for audio classification using convolutional deep belief networks. In
Proceedings of Neural Information Processing Systems (NIPS). 2010.
[218] P. Lena, K. Nagata, and P. Baldi. Deep spatiotemporal architectures
and learning for protein structure prediction. In Proceedings of Neural
Information Processing Systems (NIPS). 2012.
[219] S. Levine. Exploring deep and recurrent architectures for optimal con-
trol. arXiv:1311.1761v1.
[220] J. Li, L. Deng, Y. Gong, and R. Haeb-Umbach. An overview of
noise-robust automatic speech recognition. IEEE/Association for Com-
puting Machinery (ACM) Transactions on Audio, Speech, and Language
Processing, pages 1–33, 2014.
[221] J. Li, D. Yu, J. Huang, and Y. Gong. Improving wideband speech
recognition using mixed-bandwidth training data in CD-DNN-HMM.
In Proceedings of IEEE Spoken Language Technology (SLT). 2012.
[222] L. Li, Y. Zhao, D. Jiang, and Y. Zhang etc. Hybrid deep neural network–
hidden markov model (DNN-HMM) based speech emotion recognition.
In Proceedings Conference on Affective Computing and Intelligent Inter-
action (ACII), pages 312–317. September 2013.
[223] H. Liao. Speaker adaptation of context dependent deep neural net-
works. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2013.
368 References

[224] H. Liao, E. McDermott, and A. Senior. Large scale deep neural network
acoustic modeling with semi-supervised training data for youtube video
transcription. In Proceedings of the Automatic Speech Recognition and
Understanding Workshop (ASRU). 2013.
[225] H. Lin, L. Deng, D. Yu, Y. Gong, A. Acero, and C.-H. Lee. A study on
multilingual acoustic modeling for large vocabulary ASR. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2009.
[226] Y. Lin, F. Lv, S. Zhu, M. Yang, T. Cour, K. Yu, L. Cao, and T. Huang.
Large-scale image classification: Fast feature extraction and SVM train-
ing. In Proceedings of Computer Vision and Pattern Recognition
(CVPR). 2011.
[227] Z. Ling, L. Deng, and D. Yu. Modeling spectral envelopes using
restricted boltzmann machines and deep belief networks for statisti-
cal parametric speech synthesis. IEEE Transactions on Audio Speech
Language Processing, 21(10):2129–2139, 2013.
[228] Z. Ling, L. Deng, and D. Yu. Modeling spectral envelopes using
restricted boltzmann machines for statistical parametric speech synthe-
sis. In International Conference on Acoustics Speech and Signal Pro-
cessing (ICASSP), pages 7825–7829. 2013.
[229] Z. Ling, K. Richmond, and J. Yamagishi. Articulatory control of HMM-
based parametric speech synthesis using feature-space-switched multi-
ple regression. IEEE Transactions on Audio, Speech, and Language
Processing, 21, January 2013.
[230] L. Lu, K. Chin, A. Ghoshal, and S. Renals. Joint uncertainty decoding
for noise robust subspace gaussian mixture models. IEEE Transactions
on Audio, Speech, and Language Processing, 21(9):1791–1804, 2013.
[231] J. Ma and L. Deng. A path-stack algorithm for optimizing dynamic
regimes in a statistical hidden dynamical model of speech. Computer,
Speech and Language, 2000.
[232] J. Ma and L. Deng. Efficient decoding strategies for conversational
speech recognition using a constrained nonlinear state-space model.
IEEE Transactions on Speech and Audio Processing, 11(6):590–602,
2003.
[233] J. Ma and L. Deng. Target-directed mixture dynamic models for spon-
taneous speech recognition. IEEE Transactions on Speech and Audio
Processing, 12(1):47–58, 2004.
References 369

[234] A. Maas, A. Hannun, and A. Ng. Rectifier nonlinearities improve neural

network acoustic models. International Conference on Machine Learn-
ing (ICML) Workshop on Deep Learning for Audio, Speech, and Lan-
guage Processing, 2013.
[235] A. Maas, Q. Le, T. O’Neil, O. Vinyals, P. Nguyen, and P. Ng. Recurrent
neural networks for noise reduction in robust ASR. In Proceedings of
Interspeech. 2012.
[236] C. Manning, P. Raghavan, and H. Schütze. Introduction to Information
Retrieval. Cambridge University Press, 2009.
[237] J. Markoff. Scientists see promise in deep-learning programs. New York
Times, November 24 2012.
[238] J. Martens. Deep learning with hessian-free optimization. In Proceedings
of International Conference on Machine Learning (ICML). 2010.
[239] J. Martens and I. Sutskever. Learning recurrent neural networks with
hessian-free optimization. In Proceedings of International Conference
on Machine Learning (ICML). 2011.
[240] D. McAllester. A PAC-bayesian tutorial with a dropout bound. ArX-
ive1307.2118, July 2013.
[241] I. McGraw, I. Badr, and J. R. Glass. Learning lexicons from speech
using a pronunciation mixture model. IEEE Transactions on Audio,
Speech, and Language Processing, 21(2):357,366, February 2013.
[242] G. Mesnil, X. He, L. Deng, and Y. Bengio. Investigation of recurrent-
neural-network architectures and learning methods for spoken language
understanding. In Proceedings of Interspeech. 2013.
[243] Y. Miao and F. Metze. Improving low-resource CD-DNN-HMM using
dropout and multilingual DNN training. In Proceedings of Interspeech.
2013.
[244] Y. Miao, S. Rawat, and F. Metze. Deep maxout networks for low
resource speech recognition. In Proceedings of the Automatic Speech
Recognition and Understanding Workshop (ASRU). 2013.
[245] T. Mikolov. Statistical language models based on neural networks.
Ph.D. thesis, Brno University of Technology, 2012.
[246] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Efficient estimation of
word representations in vector space. In Proceedings of International
Conference on Learning Representations (ICLR). 2013.
370 References

[247] T. Mikolov, A. Deoras, D. Povey, L. Burget, and J. Cernocky. Strategies

for training large scale neural network language models. In Proceedings
of the IEEE Automatic Speech Recognition and Understanding Work-
shop (ASRU). 2011.
[248] T. Mikolov, M. Karafiat, L. Burget, J. Cernocky, and S. Khudanpur.
Recurrent neural network based language model. In Proceedings of
International Conference on Acoustics Speech and Signal Processing
(ICASSP), pages 1045–1048. 2010.
[249] T. Mikolov, Q. Le, and I. Sutskever. Exploiting similarities among lan-
guages for machine translation. arXiv:1309.4168v1, 2013.
[250] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean. Distributed
representations of words and phrases and their compositionality. In
Proceedings of Neural Information Processing Systems (NIPS). 2013.
[251] Y. Minami, E. McDermott, A. Nakamura, and S. Katagiri. A recogni-
tion method with parametric trajectory synthesized using direct rela-
tions between static and dynamic feature vector time series. In Pro-
ceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP), pages 957–960. 2002.
[252] A. Mnih and G. Hinton. Three new graphical models for statistical lan-
guage modeling. In Proceedings of International Conference on Machine
Learning (ICML), pages 641–648. 2007.
[253] A. Mnih and G. Hinton. A scalable hierarchical distributed lan-
guage model. In Proceedings of Neural Information Processing Systems
(NIPS), pages 1081–1088. 2008.
[254] A. Mnih and K. Kavukcuoglu. Learning word embeddings efficiently
with noise-contrastive estimation. In Proceedings of Neural Information
Processing Systems (NIPS). 2013.
[255] A. Mnih and W.-T. Teh. A fast and simple algorithm for training
neural probabilistic language models. In Proceedings of International
Conference on Machine Learning (ICML), pages 1751–1758. 2012.
[256] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wier-
stra, and M. Riedmiller. Playing arari with deep reinforcement learning.
Neural Information Processing Systems (NIPS) Deep Learning Work-
shop, 2013. also arXiv:1312.5602v1.
[257] A. Mohamed, G. Dahl, and G. Hinton. Deep belief networks for phone
recognition. In Proceedings of Neural Information Processing Systems
(NIPS) Workshop Deep Learning for Speech Recognition and Related
Applications. 2009.
References 371

[258] A. Mohamed, G. Dahl, and G. Hinton. Acoustic modeling using deep

belief networks. IEEE Transactions on Audio, Speech, & Language Pro-
cessing, 20(1), January 2012.
[259] A. Mohamed, G. Hinton, and G. Penn. Understanding how deep belief
networks perform acoustic modelling. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2012.
[260] A. Mohamed, D. Yu, and L. Deng. Investigation of full-sequence train-
ing of deep belief networks for speech recognition. In Proceedings of
Interspeech. 2010.
[261] N. Morgan. Deep and wide: Multiple layers in automatic speech recog-
nition. IEEE Transactions on Audio, Speech, & Language Processing,
20(1), January 2012.
[262] N. Morgan, Q. Zhu, A. Stolcke, K. Sonmez, S. Sivadas, T. Shinozaki,
M. Ostendorf, P. Jain, H. Hermansky, D. Ellis, G. Doddington, B. Chen,
O. Cretin, H. Bourlard, and M. Athineos. Pushing the envelope — aside
[speech recognition]. IEEE Signal Processing Magazine, 22(5):81–88,
September 2005.
[263] F. Morin and Y. Bengio. Hierarchical probabilistic neural network lan-
guage models. In Proceedings of Artificial Intelligence and Statistics
(AISTATS). 2005.
[264] K. Murphy. Machine Learning — A Probabilistic Perspective. The MIT
Press, 2012.
[265] V. Nair and G. Hinton. 3-d object recognition with deep belief nets. In
Proceedings of Neural Information Processing Systems (NIPS). 2009.
[266] T. Nakashika, R. Takashima, T. Takiguchi, and Y. Ariki. Voice conver-
sion in high-order eigen space using deep belief nets. In Proceedings of
Interspeech. 2013.
[267] H. Ney. Speech translation: Coupling of recognition and translation. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 1999.
[268] J. Ngiam, Z. Chen, P. Koh, and A. Ng. Learning deep energy models. In
Proceedings of International Conference on Machine Learning (ICML).
2011.
[269] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Ng. Multimodal
deep learning. In Proceedings of International Conference on Machine
Learning (ICML). 2011.
372 References

[270] M. Norouzi, T. Mikolov, S. Bengio, J. Shlens, A. Frome, G. Corrado,

and J. Dean. Zero-shot learning by convex combination of semantic
embeddings. arXiv:1312.5650v2, 2013.
[271] N. Oliver, A. Garg, and E. Horvitz. Layered representations for learning
and inferring office activity from multiple sensory channels. Computer
Vision and Image Understanding, 96:163–180, 2004.
[272] B. Olshausen. Can ‘deep learning’ offer deep insights about visual rep-
resentation? Neural Information Processing Systems (NIPS) Workshop
on Deep Learning and Unsupervised Feature Learning, 2012.
[273] M. Ostendorf. Moving beyond the ‘beads-on-a-string’ model of speech.
In Proceedings of the Automatic Speech Recognition and Understanding
Workshop (ASRU). 1999.
[274] M. Ostendorf, V. Digalakis, and O. Kimball. From HMMs to segment
models: A unified view of stochastic modeling for speech recognition.
IEEE Transactions on Speech and Audio Processing, 4(5), September
1996.
[275] L. Oudre, C. Fevotte, and Y. Grenier. Probabilistic template-based
chord recognition. IEEE Transactions on Audio, Speech, and Language
Processing, 19(8):2249–2259, November 2011.
[276] H. Palangi, L. Deng, and R. Ward. Learning input and recurrent weight
matrices in echo state networks. Neural Information Processing Systems
(NIPS) Deep Learning Workshop, December 2013.
[277] H. Palangi, R. Ward, and L. Deng. Using deep stacking network to
improve structured compressive sensing with multiple measurement vec-
tors. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2013.
[278] G. Papandreou, A. Katsamanis, V. Pitsikalis, and P. Maragos. Adap-
tive multimodal fusion by uncertainty compensation with application to
audiovisual speech recognition. IEEE Transactions on Audio, Speech,
and Language Processing, 17:423–435, 2009.
[279] R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio. How to construct deep
recurrent neural networks. In Proceedings of International Conference
on Learning Representations (ICLR). 2014.
[280] R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training
recurrent neural networks. In Proceedings of International Conference
on Machine Learning (ICML). 2013.
[281] J. Peng, L. Bo, and J. Xu. Conditional neural fields. In Proceedings of
Neural Information Processing Systems (NIPS). 2009.
References 373

[282] P. Picone, S. Pike, R. Regan, T. Kamm, J. bridle, L. Deng, Z. Ma,

H. Richards, and M. Schuster. Initial evaluation of hidden dynamic
models on conversational speech. In Proceedings of International Con-
ference on Acoustics Speech and Signal Processing (ICASSP). 1999.
[283] J. Pinto, S. Garimella, M. Magimai-Doss, H. Hermansky, and
H. Bourlard. Analysis of MLP-based hierarchical phone posterior prob-
ability estimators. IEEE Transactions on Audio, Speech, and Language
Processing, 19(2), February 2011.
[284] C. Plahl, T. Sainath, B. Ramabhadran, and D. Nahamoo. Improved
pre-training of deep belief networks using sparse encoding symmet-
ric machines. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2012.
[285] C. Plahl, R. Schlüter, and H. Ney. Hierarchical bottleneck features for
LVCSR. In Proceedings of Interspeech. 2010.
[286] T. Plate. Holographic reduced representations. IEEE Transactions on
Neural Networks, 6(3):623–641, May 1995.
[287] T. Poggio. How the brain might work: The role of information and
learning in understanding and replicating intelligence. In G. Jacovitt,
A. Pettorossi, R. Consolo, and V. Senni, editors, Information: Science
and Technology for the New Century, pages 45–61. Lateran University
Press, 2007.
[288] J. Pollack. Recursive distributed representations. Artificial Intelligence,
46:77–105, 1990.
[289] H. Poon and P. Domingos. Sum-product networks: A new deep archi-
tecture. In Proceedings of Uncertainty in Artificial Intelligence. 2011.
[290] D. Povey and P. Woodland. Minimum phone error and I-smoothing
for improved discriminative training. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2002.
[291] R. Prabhavalkar and E. Fosler-Lussier. Backpropagation training for
multilayer conditional random field based phone recognition. In Pro-
ceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2010.
[292] A. Prince and P. Smolensky. Optimality: From neural networks to uni-
versal grammar. Science, 275:1604–1610, 1997.
[293] L. Rabiner. A tutorial on hidden markov models and selected applica-
tions in speech recognition. In Proceedings of the IEEE, pages 257–286.
1989.
374 References

[294] M. Ranzato, Y. Boureau, and Y. LeCun. Sparse feature learning for

deep belief networks. In Proceedings of Neural Information Processing
Systems (NIPS). 2007.
[295] M. Ranzato, S. Chopra, Y. LeCun, and F.-J. Huang. Energy-based
models in document recognition and computer vision. In Proceed-
ings of International Conference on Document Analysis and Recognition
(ICDAR). 2007.
[296] M. Ranzato and G. Hinton. Modeling pixel means and covariances using
factorized third-order boltzmann machines. In Proceedings of Computer
Vision and Pattern Recognition (CVPR). 2010.
[297] M. Ranzato, C. Poultney, S. Chopra, and Y. LeCun. Efficient learning
of sparse representations with an energy-based model. In Proceedings
of Neural Information Processing Systems (NIPS). 2006.
[298] M. Ranzato, J. Susskind, V. Mnih, and G. Hinton. On deep generative
models with applications to recognition. In Proceedings of Computer
Vision and Pattern Recognition (CVPR). 2011.
[299] C. Rathinavalu and L. Deng. Construction of state-dependent dynamic
parameters by maximum likelihood: Applications to speech recognition.
Signal Processing, 55(2):149–165, 1997.
[300] S. Rennie, K. Fouset, and P. Dognin. Factorial hidden restricted boltz-
mann machines for noise robust speech recognition. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2012.
[301] S. Rennie, H. Hershey, and P. Olsen. Single-channel multi-talker speech
recognition — graphical modeling approaches. IEEE Signal Processing
Magazine, 33:66–80, 2010.
[302] M. Riedmiller and H. Braun. A direct adaptive method for faster back-
propagation learning: The RPROP algorithm. In Proceedings of the
IEEE International Conference on Neural Networks. 1993.
[303] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio. Contractive
autoencoders: Explicit invariance during feature extraction. In Proceed-
ings of International Conference on Machine Learning (ICML), pages
833–840. 2011.
[304] A. Robinson. An application of recurrent nets to phone probability
estimation. IEEE Transactions on Neural Networks, 5:298–305, 1994.
[305] T. Sainath, L. Horesh, B. Kingsbury, A. Aravkin, and B. Ramabhad-
ran. Accelerating hessian-free optimization for deep neural networks by
implicit pre-conditioning and sampling. arXiv: 1309.1508v3, 2013.
References 375

[306] T. Sainath, B. Kingsbury, A. Mohamed, G. Dahl, G. Saon, H. Soltau,

T. Beran, A. Aravkin, and B. Ramabhadran. Improvements to deep
convolutional neural networks for LVCSR. In Proceedings of the Auto-
matic Speech Recognition and Understanding Workshop (ASRU). 2013.
[307] T. Sainath, B. Kingsbury, A. Mohamed, and B. Ramabhadran. Learn-
ing filter banks within a deep neural network framework. In Proceed-
ings of The Automatic Speech Recognition and Understanding Workshop
(ASRU). 2013.
[308] T. Sainath, B. Kingsbury, and B. Ramabhadran. Autoencoder bottle-
neck features using deep belief networks. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2012.
[309] T. Sainath, B. Kingsbury, B. Ramabhadran, P. Novak, and
A. Mohamed. Making deep belief networks effective for large vocab-
ulary continuous speech recognition. In Proceedings of the Automatic
Speech Recognition and Understanding Workshop (ASRU). 2011.
[310] T. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy, and B. Ramabhad-
ran. Low-rank matrix factorization for deep neural network training
with high-dimensional output targets. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2013.
[311] T. Sainath, B. Kingsbury, H. Soltau, and B. Ramabhadran. Optimiza-
tion techniques to improve training speed of deep neural networks for
large speech tasks. IEEE Transactions on Audio, Speech, and Language
Processing, 21(11):2267–2276, November 2013.
[312] T. Sainath, A. Mohamed, B. Kingsbury, and B. Ramabhadran. Con-
volutional neural networks for LVCSR. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2013.
[313] T. Sainath, B. Ramabhadran, M. Picheny, D. Nahamoo, and
D. Kanevsky. Exemplar-based sparse representation features: From
TIMIT to LVCSR. IEEE Transactions on Speech and Audio Processing,
November 2011.
[314] R. Salakhutdinov and G. Hinton. Semantic hashing. In Proceedings of
Special Interest Group on Information Retrieval (SIGIR) Workshop on
Information Retrieval and Applications of Graphical Models. 2007.
[315] R. Salakhutdinov and G. Hinton. Deep boltzmann machines. In Pro-
ceedings of Artificial Intelligence and Statistics (AISTATS). 2009.
[316] R. Salakhutdinov and G. Hinton. A better way to pretrain deep boltz-
mann machines. In Proceedings of Neural Information Processing Sys-
tems (NIPS). 2012.
376 References

[317] G. Saon, H. Soltau, D. Nahamoo, and M. Picheny. Speaker adaptation

of neural network acoustic models using i-vectors. In Proceedings of the
Automatic Speech Recognition and Understanding Workshop (ASRU).
2013.
[318] R. Sarikaya, G. Hinton, and B. Ramabhadran. Deep belief nets for nat-
ural language call-routing. In Proceedings of International Conference
on Acoustics Speech and Signal Processing (ICASSP), pages 5680–5683.
2011.
[319] E. Schmidt and Y. Kim. Learning emotion-based acoustic features with
deep belief networks. In Proceedings IEEE of Signal Processing to Audio
and Acoustics. 2011.
[320] H. Schwenk. Continuous space translation models for phrase-based sta-
tistical machine translation. In Proceedings of Computional Linguistics.
2012.
[321] H. Schwenk, A. Rousseau, and A. Mohammed. Large, pruned or contin-
uous space language models on a gpu for statistical machine translation.
In Proceedings of the Joint Human Language Technology Conference
and the North American Chapter of the Association of Computational
Linguistics (HLT-NAACL) 2012 Workshop on the future of language
modeling for Human Language Technology (HLT), pages 11–19.
[322] F. Seide, H. Fu, J. Droppo, G. Li, and D. Yu. On parallelizability of
stochastic gradient descent for speech DNNs. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2014.
[323] F. Seide, G. Li, X. Chen, and D. Yu. Feature engineering in context-
dependent deep neural networks for conversational speech transcription.
In Proceedings of the Automatic Speech Recognition and Understanding
Workshop (ASRU), pages 24–29. 2011.
[324] F. Seide, G. Li, and D. Yu. Conversational speech transcription using
context-dependent deep neural networks. In Proceedings of Interspeech,
pages 437–440. 2011.
[325] M. Seltzer, D. Yu, and E. Wang. An investigation of deep neural net-
works for noise robust speech recognition. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP).
2013.
[326] M. Shannon, H. Zen, and W. Byrne. Autoregressive models for statisti-
cal parametric speech synthesis. IEEE Transactions on Audio, Speech,
Language Processing, 21(3):587–597, 2013.
References 377

[327] H. Sheikhzadeh and L. Deng. Waveform-based speech recognition using

hidden filter models: Parameter selection and sensitivity to power nor-
malization. IEEE Transactions on on Speech and Audio Processing
(ICASSP), 2:80–91, 1994.
[328] Y. Shen, X. He, J. Gao, L. Deng, and G. Mesnil. Learning semantic
representations using convolutional neural networks for web search. In
Proceedings World Wide Web. 2014.
[329] K. Simonyan, A. Vedaldi, and A. Zisserman. Deep fisher networks for
large-scale image classification. In Proceedings of Neural Information
Processing Systems (NIPS). 2013.
[330] M. Siniscalchi, J. Li, and C. Lee. Hermitian polynomial for speaker
adaptation of connectionist speech recognition systems. IEEE Trans-
actions on Audio, Speech, and Language Processing, 21(10):2152–2161,
2013a.
[331] M. Siniscalchi, T. Svendsen, and C.-H. Lee. A bottom-up modular
search approach to large vocabulary continuous speech recognition.
IEEE Transactions on Audio, Speech, Language Processing, 21, 2013.
[332] M. Siniscalchi, D. Yu, L. Deng, and C.-H. Lee. Exploiting deep neu-
ral networks for detection-based speech recognition. Neurocomputing,
106:148–157, 2013.
[333] M. Siniscalchi, D. Yu, L. Deng, and C.-H. Lee. Speech recognition using
long-span temporal patterns in a deep network model. IEEE Signal
Processing Letters, 20(3):201–204, March 2013.
[334] G. Sivaram and H. Hermansky. Sparse multilayer perceptrons for
phoneme recognition. IEEE Transactions on Audio, Speech, & Lan-
guage Processing, 20(1), January 2012.
[335] P. Smolensky. Tensor product variable binding and the representation
of symbolic structures in connectionist systems. Artificial Intelligence,
46:159–216, 1990.
[336] P. Smolensky and G. Legendre. The Harmonic Mind — From Neu-
ral Computation to Optimality-Theoretic Grammar. The MIT Press,
Cambridge, MA, 2006.
[337] J. Snoek, H. Larochelle, and R. Adams. Practical bayesian optimization
of machine learning algorithms. In Proceedings of Neural Information
Processing Systems (NIPS). 2012.
[338] R. Socher. New directions in deep learning: Structured models, tasks,
and datasets. Neural Information Processing Systems (NIPS) Workshop
on Deep Learning and Unsupervised Feature Learning, 2012.
378 References

[339] R. Socher, Y. Bengio, and C. Manning. Deep learning for NLP.

Tutorial at Association of Computational Logistics (ACL), 2012, and
North American Chapter of the Association of Computational Linguis-
tics (NAACL), 2013. http://www.socher.org/index.php/DeepLearning
Tutorial.
[340] R. Socher, D. Chen, C. Manning, and A. Ng. Reasoning with neural
tensor networks for knowledge base completion. In Proceedings of Neural
Information Processing Systems (NIPS). 2013.
[341] R. Socher and L. Fei-Fei. Connecting modalities: Semi-supervised
segmentation and annotation of images using unaligned text corpora.
In Proceedings of Computer Vision and Pattern Recognition (CVPR).
2010.
[342] R. Socher, M. Ganjoo, H. Sridhar, O. Bastani, C. Manning, and A. Ng.
Zero-shot learning through cross-modal transfer. In Proceedings of Neu-
ral Information Processing Systems (NIPS). 2013b.
[343] R. Socher, Q. Le, C. Manning, and A. Ng. Grounded compositional
semantics for finding and describing images with sentences. Neu-
ral Information Processing Systems (NIPS) Deep Learning Workshop,
2013c.
[344] R. Socher, C. Lin, A. Ng, and C. Manning. Parsing natural scenes
and natural language with recursive neural networks. In Proceedings of
International Conference on Machine Learning (ICML). 2011.
[345] R. Socher, J. Pennington, E. Huang, A. Ng, and C. Manning. Dynamic
pooling and unfolding recursive autoencoders for paraphrase detection.
In Proceedings of Neural Information Processing Systems (NIPS). 2011.
[346] R. Socher, J. Pennington, E. Huang, A. Ng, and C. Manning. Semi-
supervised recursive autoencoders for predicting sentiment distribu-
tions. In Proceedings of Empirical Methods in Natural Language Pro-
cessing (EMNLP). 2011.
[347] R. Socher, A. Perelygin, J. Wu, J. Chuang, C. Manning, A. Ng, and
C. Potts. Recursive deep models for semantic compositionality over a
sentiment treebank. In Proceedings of Empirical Methods in Natural
Language Processing (EMNLP). 2013.
[348] N. Srivastava and R. Salakhutdinov. Multimodal learning with deep
boltzmann machines. In Proceedings of Neural Information Processing
Systems (NIPS). 2012.
References 379

[349] N. Srivastava and R. Salakhutdinov. Discriminative transfer learning

with tree-based priors. In Proceedings of Neural Information Processing
Systems (NIPS). 2013.
[350] R. Srivastava, J. Masci, S. Kazerounian, F. Gomez, and J. Schmidhuber.
Compete to compute. In Proceedings of Neural Information Processing
Systems (NIPS). 2013.
[351] T. Stafylakis, P. Kenny, M. Senoussaoui, and P. Dumouchel. Prelimi-
nary investigation of boltzmann machine classifiers for speaker recogni-
tion. In Proceedings of Odyssey, pages 109–116. 2012.
[352] V. Stoyanov, A. Ropson, and J. Eisner. Empirical risk minimization of
graphical model parameters given approximate inference, decoding, and
model structure. In Proceedings of Artificial Intelligence and Statistics
(AISTATS). 2011.
[353] H. Su, G. Li, D. Yu, and F. Seide. Error back propagation for sequence
training of context-dependent deep networks for conversational speech
transcription. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013.
[354] A. Subramanya, L. Deng, Z. Liu, and Z. Zhang. Multi-sensory speech
processing: Incorporating automatically extracted hidden dynamic
information. In Proceedings of IEEE International Conference on Mul-
timedia & Expo (ICME). Amsterdam, July 2005.
[355] J. Sun and L. Deng. An overlapping-feature based phonological model
incorporating linguistic constraints: Applications to speech recognition.
Journal on Acoustical Society of America, 111(2):1086–1101, 2002.
[356] I. Sutskever. Training recurrent neural networks. Ph.D. Thesis, Univer-
sity of Toronto, 2013.
[357] I. Sutskever, J. Martens, and G. Hinton. Generating text with recurrent
neural networks. In Proceedings of International Conference on Machine
Learning (ICML). 2011.
[358] Y. Tang and C. Eliasmith. Deep networks for robust visual recogni-
tion. In Proceedings of International Conference on Machine Learning
(ICML). 2010.
[359] Y. Tang and R. Salakhutdinov. Learning Stochastic Feedforward Neural
Networks. NIPS, 2013.
[360] A. Tarralba, R. Fergus, and Y. Weiss. Small codes and large image
databases for recognition. In Proceedings of Computer Vision and Pat-
tern Recognition (CVPR). 2008.
380 References

[361] G. Taylor, G. E. Hinton, and S. Roweis. Modeling human motion using

binary latent variables. In Proceedings of Neural Information Processing
Systems (NIPS). 2007.
[362] S. Thomas, M. Seltzer, K. Church, and H. Hermansky. Deep neural
network features and semi-supervised training for low resource speech
recognition. In Proceedings of Interspeech. 2013.
[363] T. Tieleman. Training restricted boltzmann machines using approx-
imations to the likelihood gradient. In Proceedings of International
Conference on Machine Learning (ICML). 2008.
[364] K. Tokuda, Y. Nankaku, T. Toda, H. Zen, H. Yamagishi, and K. Oura.
Speech synthesis based on hidden markov models. Proceedings of the
IEEE, 101(5):1234–1252, 2013.
[365] F. Triefenbach, A. Jalalvand, K. Demuynck, and J.-P. Martens. Acoustic
modeling with hierarchical reservoirs. IEEE Transactions on Audio,
Speech, and Language Processing, 21(11):2439–2450, November 2013.
[366] G. Tur, L. Deng, D. Hakkani-Tür, and X. He. Towards deep under-
standing: Deep convex networks for semantic utterance classification. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2012.
[367] J. Turian, L. Ratinov, and Y. Bengio. Word representations: A simple
and general method for semi-supervised learning. In Proceedings of
Association for Computational Linguistics (ACL). 2010.
[368] Z. Tüske, M. Sundermeyer, R. Schlüter, and H. Ney. Context-dependent
MLPs for LVCSR: TANDEM, hybrid or both? In Proceedings of Inter-
speech. 2012.
[369] B. Uria, S. Renals, and K. Richmond. A deep neural network for
acoustic-articulatory speech inversion. Neural Information Processing
Systems (NIPS) Workshop on Deep Learning and Unsupervised Feature
Learning, 2011.
[370] R. van Dalen and M. Gales. Extended VTS for noise-robust speech
recognition. IEEE Transactions on Audio, Speech, and Language Pro-
cessing, 19(4):733–743, 2011.
[371] A. van den Oord, S. Dieleman, and B. Schrauwen. Deep content-based
music recommendation. In Proceedings of Neural Information Process-
ing Systems (NIPS). 2013.
[372] V. Vasilakakis, S. Cumani, and P. Laface. Speaker recognition by means
of deep belief networks. In Proceedings of Biometric Technologies in
Forensic Science. 2013.
References 381

[373] K. Vesely, A. Ghoshal, L. Burget, and D. Povey. Sequence-discriminative

training of deep neural networks. In Proceedings of Interspeech. 2013.
[374] K. Vesely, M. Hannemann, and L. Burget. Semi-supervised training of
deep neural networks. In Proceedings of the Automatic Speech Recogni-
tion and Understanding Workshop (ASRU). 2013.
[375] P. Vincent. A connection between score matching and denoising autoen-
coder. Neural Computation, 23(7):1661–1674, 2011.
[376] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P. Manzagol.
Stacked denoising autoencoders: Learning useful representations in a
deep network with a local denoising criterion. Journal of Machine
Learning Research, 11:3371–3408, 2010.
[377] O. Vinyals, Y. Jia, L. Deng, and T. Darrell. Learning with recursive
perceptual representations. In Proceedings of Neural Information Pro-
cessing Systems (NIPS). 2012.
[378] O. Vinyals and D. Povey. Krylov subspace descent for deep learning. In
Proceedings of Artificial Intelligence and Statistics (AISTATS). 2012.
[379] O. Vinyals and S. Ravuri. Comparing multilayer perceptron to deep
belief network tandem features for robust ASR. In Proceedings of
International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2011.
[380] O. Vinyals, S. Ravuri, and D. Povey. Revisiting recurrent neural net-
works for robust ASR. In Proceedings of International Conference on
Acoustics Speech and Signal Processing (ICASSP). 2012.
[381] S. Wager, S. Wang, and P. Liang. Dropout training as adaptive reg-
ularization. In Proceedings of Neural Information Processing Systems
(NIPS). 2013.
[382] A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. Lang. Phoneme
recognition using time-delay neural networks. IEEE Transactions on
Acoustical Speech, and Signal Processing, 37:328–339, 1989.
[383] G. Wang and K. Sim. Context-dependent modelling of deep neural
network using logistic regression. In Proceedings of the Automatic Speech
Recognition and Understanding Workshop (ASRU). 2013.
[384] G. Wang and K. Sim. Regression-based context-dependent modeling
of deep neural networks for speech recognition. IEEE/Association for
Computing Machinery (ACM) Transactions on Audio, Speech, and Lan-
guage Processing, 2014.
382 References

[385] D. Warde-Farley, I. Goodfellow, A. Courville, and Y. Bengi. An empir-

ical analysis of dropout in piecewise linear networks. In Proceedings of
International Conference on Learning Representations (ICLR). 2014.
[386] M. Welling, M. Rosen-Zvi, and G. Hinton. Exponential family harmo-
niums with an application to information retrieval. In Proceedings of
Neural Information Processing Systems (NIPS). 2005.
[387] C. Weng, D. Yu, M. Seltzer, and J. Droppo. Single-channel mixed speech
recognition using deep neural networks. In Proceedings of International
Conference on Acoustics Speech and Signal Processing (ICASSP). 2014.
[388] J. Weston, S. Bengio, and N. Usunier. Large scale image annotation:
Learning to rank with joint word-image embeddings. Machine Learning,
81(1):21–35, 2010.
[389] J. Weston, S. Bengio, and N. Usunier. Wsabie: Scaling up to large
vocabulary image annotation. In Proceedings of International Joint
Conference on Artificial Intelligence (IJCAI). 2011.
[390] S. Wiesler, J. Li, and J. Xue. Investigations on hessian-free optimization
for cross-entropy training of deep neural networks. In Proceedings of
Interspeech. 2013.
[391] M. Wohlmayr, M. Stark, and F. Pernkopf. A probabilistic interac-
tion model for multi-pitch tracking with factorial hidden markov model.
IEEE Transactions on Audio, Speech, and Language Processing, 19(4),
May 2011.
[392] D. Wolpert. Stacked generalization. Neural Networks, 5(2):241–259,
1992.
[393] S. J. Wright, D. Kanevsky, L. Deng, X. He, G. Heigold, and H. Li.
Optimization algorithms and applications for speech and language pro-
cessing. IEEE Transactions on Audio, Speech, and Language Processing,
21(11):2231–2243, November 2013.
[394] L. Xiao and L. Deng. A geometric perspective of large-margin training
of gaussian models. IEEE Signal Processing Magazine, 27(6):118–123,
November 2010.
[395] X. Xie and S. Seung. Equivalence of backpropagation and contrastive
hebbian learning in a layered network. Neural computation, 15:441–454,
2003.
[396] Y. Xu, J. Du, L. Dai, and C. Lee. An experimental study on speech
enhancement based on deep neural networks. IEEE Signal Processing
Letters, 21(1):65–68, 2014.
References 383

[397] J. Xue, J. Li, and Y. Gong. Restructuring of deep neural network

acoustic models with singular value decomposition. In Proceedings of
Interspeech. 2013.
[398] S. Yamin, L. Deng, Y. Wang, and A. Acero. An integrative and discrimi-
native technique for spoken utterance classification. IEEE Transactions
on Audio, Speech, and Language Processing, 16:1207–1214, 2008.
[399] Z. Yan, Q. Huo, and J. Xu. A scalable approach to using DNN-derived
features in GMM-HMM based acoustic modeling for LVCSR. In Pro-
ceedings of Interspeech. 2013.
[400] D. Yang and S. Furui. Combining a two-step CRF model and a joint
source-channel model for machine transliteration. In Proceedings of
Association for Computational Linguistics (ACL), pages 275–280. 2010.
[401] K. Yao, D. Yu, L. Deng, and Y. Gong. A fast maximum likelihood non-
linear feature transformation method for GMM-HMM speaker adapta-
tion. Neurocomputing, 2013a.
[402] K. Yao, D. Yu, F. Seide, H. Su, L. Deng, and Y. Gong. Adaptation of
context-dependent deep neural networks for automatic speech recogni-
tion. In Proceedings of International Conference on Acoustics Speech
and Signal Processing (ICASSP). 2012.
[403] K. Yao, G. Zweig, M. Hwang, Y. Shi, and D. Yu. Recurrent neural
networks for language understanding. In Proceedings of Interspeech.
2013.
[404] T. Yoshioka and T. Nakatani. Noise model transfer: Novel approach to
robustness against nonstationary noise. IEEE Transactions on Audio,
Speech, and Language Processing, 21(10):2182–2192, 2013.
[405] T. Yoshioka, A. Ragni, and M. Gales. Investigation of unsupervised
adaptation of DNN acoustic models with filter bank input. In Pro-
ceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2013.
[406] L. Younes. On the convergence of markovian stochastic algorithms with
rapidly decreasing ergodicity rates. Stochastics and Stochastic Reports,
65(3):177–228, 1999.
[407] D. Yu, X. Chen, and L. Deng. Factorized deep neural networks for adap-
tive speech recognition. International Workshop on Statistical Machine
Learning for Speech Processing, March 2012b.
[408] D. Yu, D. Deng, and S. Wang. Learning in the deep-structured con-
ditional random fields. Neural Information Processing Systems (NIPS)
2009 Workshop on Deep Learning for Speech Recognition and Related
Applications, 2009.
384 References

[409] D. Yu and L. Deng. Solving nonlinear estimation problems using splines.

IEEE Signal Processing Magazine, 26(4):86–90, July 2009.
[410] D. Yu and L. Deng. Deep-structured hidden conditional random fields
for phonetic recognition. In Proceedings of Interspeech. September 2010.
[411] D. Yu and L. Deng. Accelerated parallelizable neural networks learning
algorithms for speech recognition. In Proceedings of Interspeech. 2011.
[412] D. Yu and L. Deng. Deep learning and its applications to signal and
information processing. IEEE Signal Processing Magazine, pages 145–
154, January 2011.
[413] D. Yu and L. Deng. Efficient and effective algorithms for training single-
hidden-layer neural networks. Pattern Recognition Letters, 33:554–558,
2012.
[414] D. Yu, L. Deng, and G. E. Dahl. Roles of pre-training and fine-tuning in
context-dependent DBN-HMMs for real-world speech recognition. Neu-
ral Information Processing Systems (NIPS) 2010 Workshop on Deep
Learning and Unsupervised Feature Learning, December 2010.
[415] D. Yu, L. Deng, J. Droppo, J. Wu, Y. Gong, and A. Acero. Robust
speech recognition using cepstral minimum-mean-square-error noise
suppressor. IEEE Transactions on Audio, Speech, and Language Pro-
cessing, 16(5), July 2008.
[416] D. Yu, L. Deng, Y. Gong, and A. Acero. A novel framework and training
algorithm for variable-parameter hidden markov models. IEEE Trans-
actions on Audio, Speech and Language Processing, 17(7):1348–1360,
2009.
[417] D. Yu, L. Deng, X. He, and A. Acero. Large-margin minimum clas-
sification error training: A theoretical risk minimization perspective.
Computer Speech and Language, 22(4):415–429, October 2008.
[418] D. Yu, L. Deng, X. He, and X. Acero. Large-margin minimum classi-
fication error training for large-scale speech recognition tasks. In Pro-
ceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2007.
[419] D. Yu, L. Deng, G. Li, and F. Seide. Discriminative pretraining of deep
neural networks. U.S. Patent Filing, November 2011.
[420] D. Yu, L. Deng, P. Liu, J. Wu, Y. Gong, and A. Acero. Cross-lingual
speech recognition under runtime resource constraints. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2009b.
References 385

[421] D. Yu, L. Deng, and F. Seide. Large vocabulary speech recognition using
deep tensor neural networks. In Proceedings of Interspeech. 2012c.
[422] D. Yu, L. Deng, and F. Seide. The deep tensor neural network with
applications to large vocabulary speech recognition. IEEE Transactions
on Audio, Speech, and Language Processing, 21(2):388–396, 2013.
[423] D. Yu, J.-Y. Li, and L. Deng. Calibration of confidence measures in
speech recognition. IEEE Transactions on Audio, Speech and Language,
19:2461–2473, 2010.
[424] D. Yu, F. Seide, G. Li, and L. Deng. Exploiting sparseness in deep
neural networks for large vocabulary speech recognition. In Proceedings
of International Conference on Acoustics Speech and Signal Processing
(ICASSP). 2012.
[425] D. Yu and M. Seltzer. Improved bottleneck features using pre-trained
deep neural networks. In Proceedings of Interspeech. 2011.
[426] D. Yu, M. Seltzer, J. Li, J.-T. Huang, and F. Seide. Feature learning in
deep neural networks — studies on speech recognition. In Proceedings
of International Conference on Learning Representations (ICLR). 2013.
[427] D. Yu, S. Siniscalchi, L. Deng, and C. Lee. Boosting attribute and
phone estimation accuracies with deep neural networks for detection-
based speech recognition. In Proceedings of International Conference
on Acoustics Speech and Signal Processing (ICASSP). 2012.
[428] D. Yu, S. Wang, and L. Deng. Sequential labeling using deep-structured
conditional random fields. Journal of Selected Topics in Signal Process-
ing, 4:965–973, 2010.
[429] D. Yu, S. Wang, Z. Karam, and L. Deng. Language recognition using
deep-structured conditional random fields. In Proceedings of Interna-
tional Conference on Acoustics Speech and Signal Processing (ICASSP),
pages 5030–5033. 2010.
[430] D. Yu, K. Yao, H. Su, G. Li, and F. Seide. KL-divergence regularized
deep neural network adaptation for improved large vocabulary speech
recognition. In Proceedings of International Conference on Acoustics
Speech and Signal Processing (ICASSP). 2013.
[431] K. Yu, M. Gales, and P. Woodland. Unsupervised adaptation with dis-
criminative mapping transforms. IEEE Transactions on Audio, Speech,
and Language Processing, 17(4):714–723, 2009.
[432] K. Yu, Y. Lin, and H. Lafferty. Learning image representations from
the pixel level via hierarchical sparse coding. In Proceedings Computer
Vision and Pattern Recognition (CVPR). 2011.
386 References

[433] F. Zamora-Martínez, M. Castro-Bleda, and S. España-Boquera. Fast

evaluation of connectionist language models. International Conference
on Artificial Neural Networks, pages 144–151, 2009.
[434] M. Zeiler. Hierarchical convolutional deep learning in computer vision.
Ph.D. Thesis, New York University, January 2014.
[435] M. Zeiler and R. Fergus. Stochastic pooling for regularization of deep
convolutional neural networks. In Proceedings of International Confer-
ence on Learning Representations (ICLR). 2013.
[436] M. Zeiler and R. Fergus. Visualizing and understanding convolutional
networks. arXiv:1311.2901, pages 1–11, 2013.
[437] M. Zeiler, G. Taylor, and R. Fergus. Adaptive deconvolutional networks
for mid and high level feature learning. In Proceedings of International
Conference on Computer vision (ICCV). 2011.
[438] H. Zen, M. Gales, J. F. Nankaku, and Y. K. Tokuda. Product of
experts for statistical parametric speech synthesis. IEEE Transactions
on Audio, Speech, and Language Processing, 20(3):794–805, March 2012.
[439] H. Zen, Y. Nankaku, and K. Tokuda. Continuous stochastic feature
mapping based on trajectory HMMs. IEEE Transactions on Audio,
Speech, and Language Processings, 19(2):417–430, February 2011.
[440] H. Zen, A. Senior, and M. Schuster. Statistical parametric speech syn-
thesis using deep neural networks. In Proceedings of International Con-
ference on Acoustics Speech and Signal Processing (ICASSP), pages
7962–7966. 2013.
[441] X. Zhang, J. Trmal, D. Povey, and S. Khudanpur. Improving deep
neural network acoustic models using generalized maxout networks. In
Proceedings of International Conference on Acoustics Speech and Signal
Processing (ICASSP). 2014.
[442] X. Zhang and J. Wu. Deep belief networks based voice activity detec-
tion. IEEE Transactions on Audio, Speech, and Language Processing,
21(4):697–710, 2013.
[443] Z. Zhang, Z. Liu, M. Sinclair, A. Acero, L. Deng, J. Droppo, X. Huang,
and Y. Zheng. Multi-sensory microphones for robust speech detection,
enhancement and recognition. In Proceedings of International Confer-
ence on Acoustics Speech and Signal Processing (ICASSP). 2004.
[444] Y. Zhao and B. Juang. Nonlinear compensation using the gauss-newton
method for noise-robust speech recognition. IEEE Transactions on
Audio, Speech, and Language Processing, 20(8):2191–2206, 2012.
References 387

[445] W. Zou, R. Socher, D. Cer, and C. Manning. Bilingual word embed-

dings for phrase-based machine translation. In Proceedings of Empirical
Methods in Natural Language Processing (EMNLP). 2013.
[446] G. Zweig and P. Nguyen. A segmental CRF approach to large vocab-
ulary continuous speech recognition. In Proceedings of the Automatic
Speech Recognition and Understanding Workshop (ASRU). 2009.