Introduction to Deep Learning

Radu Ionescu, Prof. PhD.

[email protected]

Faculty of Mathematics and Computer Science

University of Bucharest
 What is this class about?

• Some of the most exciting

developments in …

• Machine Learning, Vision, NLP,

Speech, Robotics & AI in general

• … in the last decade!

 Instructors
• Lectures:
 Radu Ionescu ([email protected])

• Labs:
 Mihaela Găman ([email protected])
 Antonio Bărbălău ([email protected])
 Prerequisites
• Practical Machine Learning
– Classifiers, regressors, loss functions, normalization, MLE,
etc.

• Linear Algebra
– Matrix multiplication, eigenvalues, etc.

• Calculus
– Multi-variate gradients, hessians, jacobians, etc.

• Programming!
– Projects will require Python
– Libraries/Frameworks: Numpy, OpenCV, TensorFlow /
PyTorch
 Grading System
 Your final is based on 1 or 2 projects:

 Project 1 based on some vision classification/regression task

 Project 2 based on some NLP classification/regression task

 Both projects are individual! (NO collaboration allowed)

 Only one project is mandatory, but you can try both!

 Projects must be presented no later than the day of the “exam”

 There will be no other exam!

 Grading System
 Each project consists of implementing some deep learning
method(s) for the proposed Kaggle challenge (TBA)
 The grades will be proportional to your accuracy:
- Top 1-5 => your grade can be up to 10
- Top 6-10 => your grade can be up to 9
- Top 11-15 => your grade can be up to 8
- Top 16-20 => your grade can be up to 7
- Top 21-25 => your grade can be up to 6
- Others => your grade can be up to 5
• Participants that rank in the same range in both challenges, get
an extra bonus point
 Grading System
 For a grade higher or equal to 5, you must beat the baseline!

 Project(s) must be presented (2 points awarded for presentation

and documentation)

 The project consists of the code implementation in Python (any

library is allowed) and a report/documentation including:

 a description of the implemented deep learning method(s):

architecture, hyperparameters, loss, etc.
 figures and/or tables with results (including validation,
hyperparameter tuning, grid search, random search)
 comments on the results
 conclusion
 (NO) Collaboration Policy

• Collaboration
– Each student must write their own code for the project(s)

• No tolerance on plagiarism
– Neither ethical nor in your best interest
– Don’t cheat. We will find out (code will be checked!)
Acquisitions
 Topics in Practical ML
• Basics of Statistical Learning
• Loss function, MLE, MAP, Bayesian estimation, bias-variance tradeoff,
overfitting, regularization, cross-validation

• Supervised Learning
• Nearest Neighbour, Naïve Bayes, Logistic Regression, Support Vector
Machines, Kernels, Neural Networks, Decision Trees
• Ensemble Methods

• Unsupervised Learning
• Clustering: k-means, Gaussian mixture models, EM
• Dimensionality reduction: PCA, SVD, LDA

• Perception
• Applications to Vision, Natural Language Processing
 What is Machine Learning?
• “the acquisition of knowledge or skills
through experience, study, or by being
taught”
 What is Machine Learning?
• [Arthur Samuel, 1959]
– Field of study that gives computers
– the ability to learn without being explicitly programmed

• [Kevin Murphy] algorithms that

– automatically detect patterns in data
– use the uncovered patterns to predict future data or other
outcomes of interest

• [Tom Mitchell] algorithms that

– improve their performance (P)
– at some task (T)
– with experience (E)
 What is Machine Learning?

Machine
Data Understanding
Learning
 ML in a Nutshell
• Tens of thousands of machine learning
algorithms
– Hundreds new every year

• Decades of ML research oversimplified:

– All of Machine Learning:
– Learn a mapping from input to output f: X  Y
• e.g. X: emails, Y: {spam, not-spam}
 Types of Learning
• Supervised learning
– Training data includes desired outputs

• Unsupervised learning
– Training data does not include desired outputs

• Weakly or Semi-supervised learning

– Training data includes a few desired outputs

• Reinforcement learning
– Rewards from sequence of actions
 Tasks
Supervised Learning
x Classification y Discrete

x Regression y Continuous

Unsupervised Learning

x Clustering y Discrete ID

x Dimensionality y Continuous
Reduction
 Supervised Learning

x Classification y Discrete
Vision: Image Classification
x y

man
camel
carrot
car
NLP: Machine Translation
Speech: Speech2Text
 AI: Turing Test
“Can machines think”

Q: Please write me a sonnet on the subject of the Forth Bridge.

A: Count me out on this one. I never could write poetry.
Q: Add 34957 to 70764.
A: (Pause about 30 seconds and then give as answer) 105621.
 AI: Visual Turing Test

Q: How many slices

x of pizza are there?

y A: 6
 Supervised Learning
• Input: x (images, text, emails…)

• Output: y (spam or non-spam…)

• (Unknown) Target Function

– f: X  Y (the “true” mapping / reality)

• Data
– (x1,y1), (x2,y2), …, (xN,yN)

• Model / Hypothesis Class

– g: X  Y
– y = g(x) = sign(wTx)

• Learning = Search in hypothesis space

– Find best g in model class.
 Synonyms
• Representation Learning

• Deep (Machine) Learning

• Deep Neural Networks

• Deep Unsupervised Learning

• Simply: Deep Learning

So what is Deep (Machine) Learning?

• A few different ideas:

• (Hierarchical) Compositionality
– Cascade of non-linear transformations
– Multiple layers of representations

• End-to-End Learning
– Learning (goal-driven) representations
– Learning to extract features

• Distributed Representations
– No single neuron “encodes” everything
– Groups of neurons work together
 Traditional Machine Learning
VISION
hand-crafted
features your favorite

SIFT/HOG
classifier “car”
fixed learned

SPEECH
hand-crafted
features your favorite

MFCC
classifier \ˈd ē p\
fixed learned

NLP
hand-crafted
This burrito place features your favorite
is yummy and fun! bag-of-words
classifier “+”
fixed learned
 It’s an old paradigm
• The first learning machine:

Feature Extractor
the Perceptron
 Built at Cornell in 1960
• The Perceptron was a linear classifier on
top of a simple feature extractor A Wi
• The vast majority of practical applications
of ML today use glorified linear classifiers
𝑁
or glorified template matching
• Designing a feature extractor requires 𝑦 = 𝑠𝑖𝑔𝑛 ෍ 𝑊𝑖 𝐹𝑖 (𝑋) + 𝑏
considerable efforts by experts 𝑖=1
 Hierarchical Compositionality
VISION

pixels edge texton motif part object

SPEECH
sample spectral formant motif phone word
band

NLP
character word NP/VP/.. clause sentence story
 Building a Complicated Function
Given a library of simple functions

Compose into a

complicated function
 Building A Complicated Function
Given a library of simple functions

Idea 1: Linear Combinations

Compose into a
• Boosting
• Kernels
complicate function
• …

+
 Building A Complicated Function
Given a library of simple functions

Idea 2: Compositions
Compose into a
• Deep Learning
• Grammar models
complicate function
• Scattering transforms…
 Building A Complicated Function
Given a library of simple functions

Idea 2: Compositions
Compose into a
• Deep Learning
• Grammar models
complicate function
• Scattering transforms…
Deep Learning = Hierarchical Compositionality

“car”
Deep Learning = Hierarchical Compositionality

Low-Level Mid-Level High-Level Trainable “car”

Feature Feature Feature Classifier

Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013]
 Sparse DBNs
[Lee et al. ICML ‘09]
Figure courtesy: Quoc Le
The Mammalian Visual Cortex is Hierarchical
• The ventral (recognition) pathway in the visual cortex

[picture from Simon Thorpe]

So what is Deep (Machine) Learning?

• A few different ideas:

• (Hierarchical) Compositionality
– Cascade of non-linear transformations
– Multiple layers of representations

• End-to-End Learning
– Learning (goal-driven) representations
– Learning to extract features

• Distributed Representations
– No single neuron “encodes” everything
– Groups of neurons work together
 Traditional Machine Learning
VISION
hand-crafted
features your favorite

SIFT/HOG
classifier “car”
fixed learned

SPEECH
hand-crafted
features your favorite

MFCC
classifier \ˈd ē p\
fixed learned

NLP
hand-crafted
This burrito place features your favorite
is yummy and fun! Bag-of-words
classifier “+”
fixed learned
Feature Engineering

SIFT Spin Images

HoG Textons

and many many more….

 What are the current bottlenecks?
• Ablation studies on DPM [Parikh & Zitnick,
CVPR10]
– Replace every “part” in the model with a
human
• Key takeaway: “parts” or features are the most
important!
 Seeing is worse than believing
• [Barbu et al. ECCV14]
Traditional Machine Learning (more accurately)
“Learned”
VISION
K-Means/
SIFT/HOG
pooling
classifier
“car”
fixed unsupervised supervised

SPEECH
Mixture of
MFCC
Gaussians
classifier
\ˈd ē p\
fixed unsupervised supervised

NLP
This burrito place Parse Tree

is yummy and fun! Syntactic

n-grams classifier
“+”
fixed unsupervised supervised
Deep Learning = End-to-End Learning
“Learned”
VISION
K-Means/
SIFT/HOG
pooling
classifier
“car”
fixed unsupervised supervised

SPEECH
Mixture of
MFCC
Gaussians
classifier
\ˈd ē p\
fixed unsupervised supervised

NLP
This burrito place Parse Tree

is yummy and fun! Syntactic

n-grams classifier
“+”
fixed unsupervised supervised
Deep Learning = End-to-End Learning
• A hierarchy of trainable feature transforms
– Each module transforms its input representation
into a higher-level one.
– High-level features are more global and more
invariant
– Low-level features are shared among categories

Trainable Trainable Trainable

Feature- Feature- Feature-
Transform / Transform / Transform /
Classifier Classifier Classifier

Learned Internal Representations

 “Shallow” vs Deep Learning

• “Shallow” models
hand-crafted “Simple” Trainable
Feature Extractor Classifier
fixed learned

• Deep models
Trainable Trainable Trainable
Feature- Feature- Feature-
Transform / Transform / Transform /
Classifier Classifier Classifier

Learned Internal Representations

Do we really need deep models?
So what is Deep (Machine) Learning?

• A few different ideas:

• (Hierarchical) Compositionality
– Cascade of non-linear transformations
– Multiple layers of representations

• End-to-End Learning
– Learning (goal-driven) representations
– Learning to extract features

• Distributed Representations
– No single neuron “encodes” everything
– Groups of neurons work together
 Distributed Representations Toy
Example
• Local vs Distributed
 Distributed Representations Toy
Example
• Can we interpret each dimension?
 Power of distributed representations!

Local

Distributed
Power of distributed representations!

• United States:Dollar :: Romania:?

Romania
Power of distributed representations!

• Example: all face images of a person

– 1000x1000 pixels = 1,000,000 dimensions
– But the face has 3 Cartesian coordinates and 3 Euler angles
– And humans have less than about 50 muscles in the face
– Hence the manifold of face images for a person has <56 dimensions
• The perfect representations of a face image:
– Its coordinates on the face manifold
– Its coordinates away from the manifold

1.2 Face/not face

Ideal
Feature
Extractor []−3
0.2
− 2 .. .
Pose
Lighting
Expression
-----
Power of distributed representations!
The Ideal Disentangling Feature Extractor

View
Pixel n
Ideal
Feature
Extractor

Pixel 2

Expression
Pixel 1
 Distributed Representations
• Q: What objects are in the image? Where?
Power of distributed representations!
So what is Deep (Machine) Learning?

• A few different ideas:

• (Hierarchical) Compositionality
– Cascade of non-linear transformations
– Multiple layers of representations

• End-to-End Learning
– Learning (goal-driven) representations
– Learning to feature extraction

• Distributed Representations
– No single neuron “encodes” everything
– Groups of neurons work together
 Benefits of Deep/Representation
Learning
• (Usually) Better Performance
– “Because gradient descent is better than you”
Yann LeCun

• New domains without “experts”

– RGBD
– Multi-spectral data
– Gene-expression data
– Unclear how to hand-engineer
“Expert” intuitions can be misleading

• “Every time I fire a linguist, the performance of our

speech recognition system goes up”
– Fred Jelinik, IBM ’98

• “Maybe the molecule didn’t go to graduate school”

– Will Welch defending the success of his approximate
molecular screening algorithm, given that he’s a
computer scientist, not a chemist
Database Screening for HIV Protease Ligands: The Influence of Binding-Site
Conformation and Representation on Ligand Selectivity", Volker Schnecke,
Leslie A. Kuhn, Proceedings of the Seventh International Conference on
Intelligent Systems for Molecular Biology, Pages 242-251, AAAI Press, 1999.
 Problems with Deep Learning
• Problem#1: Non-Convex! Non-Convex! Non-Convex!
– Depth>=3: most losses non-convex in parameters
– Theoretically, all bets are off
– Leads to stochasticity
• different initializations  different local minima

• Standard response #1
– “Yes, but all interesting learning problems are non-convex”
– For example, human learning
• Order matters  wave hands  non-convexity

• Standard response #2
– “Yes, but it often works!”
 Problems with Deep Learning
• Problem#2: Hard to track down what’s failing
– Pipeline systems have “oracle” performances at
each step
– In end-to-end systems, it’s hard to know why
things are not working
 Problems with Deep Learning
• Problem#2: Hard to track down what’s failing

[Fang et al. CVPR15] [Vinyals et al. CVPR15]

Pipeline End-to-End
 Problems with Deep Learning
• Problem#2: Hard to track down what’s failing
– Pipeline systems have “oracle” performances at each step
– In end-to-end systems, it’s hard to know why things are not
working

• Standard response #1
– Tricks of the trade: visualize features, add losses at
different layers, pre-train to avoid degenerate
initializations…
– “We’re working on it”

• Standard response #2
– “Yes, but it often works!”
 Problems with Deep Learning
• Problem#2: Hard to track down what’s failing
– There are methods to
visualize the features, e.g.:
GradCAM
GradCAM++
 Problems with Deep Learning
• Problem#3: Lack of easy reproducibility
– Direct consequence of stochasticity & non-convexity

• Standard response #1
– It’s getting much better
– Standard toolkits/libraries/frameworks now available
– TensorFlow, PyTorch, Caffe, Theano

• Standard response #2
– “Yes, but it often works!”
Yes it works, but how?
 ImageNet Large Scale Visual
Recognition Challenge (ILSVRC)
Classification:
1000 object classes 1.4M/50k/100k images
Detection:
200 object classes 400k/20k/40k images
Dalmatian

http://image-net.org/challenges/LSVRC/{2010,…,2014}
 Data Enabling Richer Models
• [Krizhevsky et al. NIPS12]
– 54 million parameters; 8 layers (5 conv, 3 fully-
connected)
– Trained on 1.4M images in ImageNet
– Better Regularization (Dropout)
1k output
Input Image
units

Convolution Layer Pooling Layer Convolution Layer Pooling Layer Fully-Connected MLP
+ Non-Linearity + Non-Linearity
 ImageNet Classification 2012
• [Krizhevsky et al. NIPS12]: 15.4% error
• Next best team: 26.2% error

(C) Dhruv Batra 77

 Other Domains & Applications
• Vision • Medical Imaging
• Natural Language Processing • Retail
• Speech • Surveillance
• Robotics • Insurance
• Game Playing • Many others
 Why are things working today?
• More compute power
– GPUs are ~50x faster

• More data
– 108 samples (compared to 103 in 1990s)

• Better algorithms/models/regularizers
– Dropout
– ReLU
– Batch-Normalization
– …
 THE SPACE OF
MACHINE LEARNING METHODS
 Recurrent
Boosting
Neural Net
Convolutional
Neural Net
Perceptron
Neural Net

SVM

Deep (sparse/denoising)
Autoencoder
Autoencoder
Sparse Coding
GMM
Deep Belief Net
Restricted BM
BayesNP
Disclaimer: showing only a
subset of the known methods
 SHALLOW
Recurrent
Boosting
Neural Net
Convolutional
Neural Net
Perceptron
Neural Net

SVM

Deep (sparse/denoising)
Autoencoder
Autoencoder
Sparse Coding
GMM
Deep Belief Net
Restricted BM
DEEP

BayesNP
 SHALLOW
Recurrent
Boosting
Neural Net
Convolutional
Neural Net
Perceptron
Neural Net

SVM
SUPERVISED
Deep (sparse/denoising) UNSUPERVISED
Autoencoder
Autoencoder
Sparse Coding
GMM
Deep Belief Net
Restricted BM
DEEP

BayesNP
 SHALLOW
Recurrent
Boosting
Neural Net
Convolutional
Neural Net
Perceptron
Neural Net

SVM
SUPERVISED
Deep (sparse/denoising) UNSUPERVISED
Autoencoder
Autoencoder
PROBABILISTIC Sparse Coding
GMM
Deep Belief Net
Restricted BM
DEEP

BayesNP
feed-forward Main types of deep architectures

Feed-back
• Neural Nets • Hierar. Sparse Coding
• Conv Nets • Deconv Nets

input input
Bi-directional

• Stacked Auto-encoders Recurrent • Recurrent Neural Nets

• BiLSTM • Recursive Nets
• LSTM

input
input
feed-forward Focus of this class

Feed-back
• Neural Nets • Hierar. Sparse Coding
• Conv Nets • Deconv Nets

input input
Bi-directional

• Stacked Auto-encoders Recurrent • Recurrent Neural Nets

• BiLSTM • Recursive Nets
• LSTM

input
input
Main types of learning protocols
• Purely supervised
• Backprop + SGD
– Good when there is lots of labeled data.

• Layer-wise unsupervised + supervised linear classifier

• Train each layer in sequence using regularized auto-encoders or RBMs
• Hold fix the feature extractor, train linear classifier on features
– Good when labeled data is scarce but there is lots of unlabeled
data.

• Layer-wise unsupervised + supervised backprop

• Train each layer in sequence
• Backprop through the whole system
– Good when learning problem is very difficult.
 Focus of this class
• Purely supervised
• Backprop + SGD
– Good when there is lots of labeled data.

• Layer-wise unsupervised + supervised linear classifier

• Train each layer in sequence using regularized auto-encoders or RBMs
• Hold fix the feature extractor, train linear classifier on features
– Good when labeled data is scarce but there is lots of unlabeled
data.

• Layer-wise unsupervised + supervised backprop

• Train each layer in sequence
• Backprop through the whole system
– Good when learning problem is very difficult.