Advance Topics in

Machine Learning

1
 Contents

• Introduction to deep learning,

• overview of reinforcement learning,
• Case study of ML applications:
– Image recognition,
– Speech recognition,
– Email spam filtering,
– Online fraud detection
– and other.

2
 AI, ML and DL

• Artificial Intelligence
– Broad field of study dedicated to
complex problem solving
• Machine learning
– Subfield of AI
– Data driven approach focused on creating algorithms that has
ability to learn from data without being explicitly programmed.
• Deep learning
– Subfield of ML focused on Deep NN
– Automatically learns hierarchical representation
– Factors behind success of DL:
• Availability of Big data
• Faster computations (GPU)
• Larger storage
3 • Algorithms
• Tools / community
 What is Deep Learning?
• There is no single definition of deep learning, but most
definitions emphasize:
– Subfield of machine learning
– Set of algorithms that attempt to model high level abstraction (key concepts)
of data
– Deep learning is just very big neural networks on a lot more data, requiring
bigger computers (faster computations, larger storage..)
– Models are graph structures (networks) with many layers (So, deep!)
– Layers in the model are combination of linear and nonlinear transformations
– Both supervised and unsupervised methods are used for fitting models to data

• Also known as
– deep structured learning or
– hierarchical learning or
– deep machine learning
4 • Deep learning has been characterized as a buzzword,
or a rebranding of neural networks.
 Introduction to Deep Learning
• Neural networks, as we have already seen in this chapter, are a class of machine
learning algorithms.
• As we have also seen, there are multiple choices of architectures for neural
networks, multi-layer neural network being one of the most adopted ones.
• However, in a multi-layer neural network, as we keep increasing the number of
hidden layers, the computation becomes very expensive.
• Going beyond two to three layers becomes quite difficult computationally.
• The only way to handle such intense computation is by using graphics processing
unit (GPU) computing.
• When we have less number of hidden layers – at the maximum two to three
layers, it is a normal neural network, which is sometimes given the fancy name
‘shallow neural network’.
• However, when the number of layers increases, it is termed as deep neural
network.
• One of the earliest deep neural networks had three hidden layers. Deep learning is
a more contemporary branding of deep neural networks, i.e. multilayer neural
networks having more than three layers.
5
• More detailed understanding of deep learning is beyond the scope of this book.
 NN History

6
 Motivation behind DL
Availability of:
Faster computer and highly optimized hardware (GPU)
Large datasets (Big data)
New models and algorithms

Better Performances on large datasets:

Performance

Traditional ML
algorithms

Size of Data

DL has capability to deal with Big data (amount and diversity – image,
7 text, voice, etc.)
Remarkable Milestone applications
 ML vs DL: A Major Difference

ML DL
• Manual feature extraction • Automatic feature
extraction
• Requires data • Requires huge data
8 • Can be performed with • Require support of
CPU processing with GPU
 Deep Learning (Yet an Another Example!)
• Deep learning has an inbuilt automatic multi stage feature
learning process that learns rich hierarchical representations (i.e.
features).
Output
Low Mid High
Trainable (e.g.
Level Level Level
Classifier outdoor,
Features Features Features
indoor)

• Image Pixel  Edge  Texture  Motif  Part  Object

• Text  Character  Word  Word-group  Clause  Sentence  Story
9
• Each module in Deep Learning transforms its input representation into a higher-
level one, in a way similar to human cortex.
 Various Deep Neural Network Architectures

Family of Methods:
• Deep Neural Networks (DNN)
• Deep Convolutional Neural Networks (Deep CNN)
• Neural History Compressor
• Recursive Neural Networks
• Long Short-Term Memory (LSTM)
• Deep Belief Networks (DBN)
• Convolution Deep Belief Networks
• Large memory storage and retrieval neural networks
• Deep Boltzman Machines (DBM)
• Generative Adversial Network (GAN)
• Stacked (de-noising) auto-encoders
• Deep Stacking networks
10 • Networks with separate memory structures
• Deep SVM
• …
 Deep Learning Application Domains

• Successful application domains

– Image: object recognition/classification and localization, character
recognition (optical & Handwritten), biometric recognition (face, finger,
gesture, etc.), expression identification
– Audio: speech recognition, speaker identification, music retrieval
– Video: Event detection, group detection, group activity detection, …
– Text (NLP): parsing, sentiment analysis, machine translation
– Bioinformatics & Biomedical informatics: Drug Discovery and
Toxicology
– Market: Customer Relationship Management, Recommendation
Systems
– Autonomous Vehicles & Drawn
– Robotics
11 – Data Analytics
– …
 An another benefit of DL Models

• Pre-trained models

• Transfer learning

12
 A Benchmark Dataset: IMAGENET

22K categories and 14M

images

• Animal • Plant • Structure • Perso

s • Bird s • Tree s n
• Fish • Flowe • Artifact
• Tools • Scene
• Indoor
• Mammal r • Appliance s • Geological
• Invertebrat • Food s • SportFormations
e • Materials • Structure Activities
s

13

https://www.kaggle.com/c/imagenet-object-localization-challenge
 Error rate on
ImageNet Dataset

14

N C Chauhan @ Department of Information Technology, A. D. Patel Instute of Technology, Anand, Gujarat, India
 Other Applications of Deep Learning:
Top-20 Applications of DL

Trend Prediction Information retrieval (search Language Translation

engines)
Robot Control Speech Recognition
Pattern recognition
Time series Generating Handwriting
Audience targeting
prediction
Face Recognition
Sentiment analysis (based on
Rhythm learning
written text) Autonomous Driving
Music composition Personalization Generating Arts
Grammar learning Automation
Imitating Famous
Handwriting Natural Language Processing Painters
recognition
Social media mining Generating Music
Human action Organic search and content
recognition Generating Photos
performance
Sign language Brand and product
15 translation differentiation
 Popular Deep Learning Frameworks

16
 Why GPU matters in DL?

17
 Reinforcement Learning

18
 Reinforcement Learning
• Reinforcement learning – A machine learns to act on its own to achieve the given goals.
• Examples: self-driving cars, intelligent robots, etc.
• We have seen babies learn to walk without any prior knowledge of how to do it. Often we
wonder how they really do it. They do it in a relatively simple way.
• First they notice somebody else walking around, for example parents or anyone living
around. They understand that legs have to be used, one at a time, to take a step. While
walking, sometimes they fall down hitting an obstacle, whereas other times they are able to
walk smoothly avoiding bumpy obstacles. When they are able to walk overcoming the
obstacle, their parents are elated and appreciate the baby with loud claps / or may be a
chocolates. When they fall down while circumventing an obstacle, obviously their parents do
not give claps or chocolates. Slowly a time comes when the babies learn from mistakes and
are able to walk with much ease.
• In the same way, machines often learn to do tasks autonomously. Let’s try to understand in
context of the example of the child learning to walk. The action tried to be achieved is
walking, the child is the agent and the place with hurdles on which the child is trying to walk
resembles the environment. It tries to improve its performance of doing the task. When a
sub-task is accomplished successfully, a reward is given. When a sub-task is not executed
correctly, obviously no reward is given. This continues till the machine is able to complete
execution of the whole task. This process of learning is known as reinforcement learning.
19 Figure 1.10 captures the high-level process of reinforcement learning.
 Reinforcement Learning
• One contemporary example of reinforcement learning is self-driving cars.
The critical information which it needs to take care of are speed and speed
limit in different road segments, traffic conditions, road conditions, weather
conditions, etc. The tasks that have to be taken care of are start/stop,
accelerate/decelerate, turn to left / right, etc.
• Popular Reinforcement Learning algorithm: Q-Learning Algorithm

20
 Applications of ML

21
 ML applications

• Image recognition
• Speech recognition
• Email spam filtering
• Online fraud detection
• Handwritten Character Recognition
• Expression Recognition
• Stock market prediction
• Recommendation systems

22