# Machine Learning Workshop

**Anthill Inside 2017**

Shubham Dokania

@shubhamdokania

## About Me

- Recent Graduate from Delhi Technological University
- B.Tech in Mathematics and Computing Engineering

- Machine Learning Instructor at Coding Blocks.
- Research Intern at IIIT-D
- Supervised by Dr. Ganesh Bagler
- Working on drug-side-efffect predictions

- A few papers (accepted and under-review) on Evolutionary Optimization, Vision and ML.
- Independent Study on Deep Learning, Reinforcement Learning.

## Workshop Overview

- Introduction to Machine Learning
- Four Sessions (~90 mins each)
- Mathematical Conceptualisation
- Project Implementation

## Session Information

- Introduction and warm-up
- CNN architectures and applications
- Transfer Learning & Sequence Processing
- Reinforcement Learning

## Session 01: Introduction

- History and Introduction to DL
- Applications (Industrial and daily)
- Convex Optimization
- Hands-on with code

## Deep Learning

- Introduction and some History
- The Gartner Hype Cycle
- Applications
- Image classification/recognnition
- Image captioning/VQA
- Sequence processing
- Speech Recognition

# Neural Networks

## A brief discussion

# Foundation

## Perceptron model

# Relation with Logistic Regression

# Neurons and ANN

## Artificial Neural Network

# Interactive Example

## Tensorflow Playground

# Gradient Descent

## Optimization

# Backpropagation

## (Chain rule for neural nets)

# What's going on inside the black box?

# Neural Nets as Feature extractors

# It's all about Manifolds

# Every layer is a transformation operation

Some examples of transformation operations

# Build your own neural network

Libraries: Numpy, Matplotlib

# Some Popular Frameworks

- Caffe
- Torch
- PyTorch
- Chainer
- Lasagne
- Tensorflow
- Theano
- Keras
- And more...

# Small Break

~10-15 mins

# Session 02:Convolutions

## And CNN

# What is Convolution?

# Some parameters...

Padding, Strides etc...

# Application to Neural Networks

# Advantages and Disadvantages over ANN

# Hands-on with CNNs

# Toy example

## The MNIST dataset

# How it works?

## Some visualizations of filters

# ImageNet datasets

## models like VGG, AlexNet, Inception-v3 etc.

# Using Pre-trained models

## Use VGG-19 from Keras

# Small Break

## Session 03: Transfer Learning

- What is it?
- Applications

# Small Demo with MNIST

# Knowledge transfer

## with VGG-16/19 Net

## Results and Discussion

- Accuracy and Loss
- Training Time
- Improvements

# Exciting Applications!

## Deep Dream and Style Transfer

Let's build both!

# Sequence Processing

## Introduction to Recurrent Architectures

# The Recurrent Neural Network

# Training of RNN

## Backprop through time

# Improvements over RNN

## Long Short-Term Memory (LSTM)

# Encoder-Decoder models

## and thought-vector representations

# Hands-On

## Music generation with LSTMs

## Applications of Previous Concepts

- AutoEncoder models
- Sequence to sequence learning
- Neural Machine Translation
- Sentiment classification

- And more...

# Small Break

## ~10-15 mins

## Session 04: Reinforcement Learning

- Supervised, Unsupervised and RL
- Learning through Rewards
- The algorithm powering AlphaGo

## Q-Learning

- A Dynamic Programming approach
- Update function
- Toy task with simple graph

## Approximate Q-Learning

- Feature based learning
- Approximate rewards
- Basis for Deep Q-Networks

# Deep Q-Network

## and solving Atari games

## Problem Statement

- Game-play understanding
- The Reward function
- DQN on Atari Ms-Pacman
- Hands-on project

# Final Project

## Train Flappy bird with RL

## Final Wrap-up

- Conclusions
- Some more examples
- What to do after the workshop
- Resources and papers to follow

# Thank You!

#### Deep Learning Applications: A hands-on approach

By Shubham Dokania

# Deep Learning Applications: A hands-on approach

Presentation for Anthill Inside 2017 workshop

- 963