Deep Learning – Spring 2017

Course Overview

We are going to take “get your hands dirty” approach, you will be given assignments and projects to implement ideas discussed in the class. Projects and assignments will contain miniature versions of real life existing applications and problems (e.g can you train your computer to generate dialogues in Shakespeare style or convert your image into painting as done by Monet, sentiment analysis etc.. ).
Course will concentrate in developing both mathematical knowledge and implementation capabilities. We will start from training a single perceptron, move to training a deep neural network, study why training large networks is a problem and what could be its possible solutions. After dipping our toes in deep belief networks and recurrent neural network we will start looking into applications of deep learning in three different areas, text-analysis, speech processing and computer vision. Objective of this approach is to make you comfortable enough that you can understand various research problems and, if interested, can implement deep learning based applications.

Course Objectives

In last few years machine learning has matured from the science fiction to reality. We are living in a world where we have already seen industry bringing to reality self-driving cars, face-recognizers that work on massive scale (facebook), speech translation systems that can translate from one language to many other simultaneously and in real-time, and more interestingly we have machines that can learn to play atari games in a similar fashion like we do.
A lot of these victories have come from the exciting field of Deep Learning; a learning methodology based on the concept that human mind captures details at multiple levels or at multiple abstract levels. One property of deep learning is removing the responsibility of humans to design features, instead Deep Learning is given a task to find the appropriate representation.

Office Hours and Contact Info.

Instructor: Dr. Mohsen Ali

2:00 pm to 4:00 pm Tuesday & Thursday (or by appointment)

Email: mohsen.ali@itu.edu.pk

Teaching Assistant: : Ali Hassan (TA Hours: Fridays 9:30 am — 11:30am)

Email: mscs15049@itu.edu.pk

Teaching Assistant: : Afsheen Rafaqat Ali (volunteer)

Email: afsheen.ali@itu.edu.pk

Prerequisites

Probability, linear algebra, programming ability and desire to read & implement.

Grading Policy

  • 50% Assignments
  • 15% Final Project
  • 10% Quizzes
  • 10% Midterm Exam
  • 15% Final Exam

Honor Code

All cases of academic misconduct will be forwarded to the disciplinary committee. All assignments are individual unless explicitly specified by the instructor. In the words of Efros, let’s not embarrass ourselves.

Course Content

Lecture Topic Assignment
Feb 2, 2017 Introduction to Deep Learning
Supervised Machine Learning
Feb 3, 2017 Linear Regression
Least Squares Estimation
Feb 7, 2017 Linear Classifier
Perceptron Rule
Sigmoid Perceptron
Reading:
Feb 9, 2017 Logistic Regression
Concepts of local minima, local maxima, convex functions, concave functions, critical points
Error Function of Linear Regression and convexity
A simple code for the Perceptron
Feb 14, 2017 Gradient descent:

Neural Networks
Feb 16, 2017 Back Propagation-1
Reading:

Feb 21, 2017 Back Propagation-2
Feb 28, 2017 How to Train Neural Network
Reading:
Mar 2, 2017 Representational or Expressive Power of Deep Neural Network (Section 6.4 Deep Learing, by Good Fellow, Bengio & Courville)
Universal Approximation Properties of Single Layer Neural Network
Problems in training, generalization and efficiency
Does increasing depth always results in much better generalization?
Convolutional Neural Networks
Mar 7, 2017 Convolutional Neural Networks
Video Lecture :

Reading :

Mar 9, 2017 CNN, Regularizations and Updation Rules
Reading: Things to Take Care of While implementing DNN

Mar 14, 2017 Network Visualization
Object Detection using Deep Learning
Apr 4, 2017 Object Localization and Detection
Regression for the localization
Concept of object proposals
RCNN/Fast-RCNN/Faster-RCNN
YOLO
Apr 6, 2017 Object Localization
Reading & Resources
Recurrent Neural Networks
Apr 11, 2017 RNN
LSTM
Image Caption Generation using LSTM
Segmentation
Apr 13, 2017 Image Segmentation Architectures
Concepts of Dropout, connect it to visualization
Apr 18, 2017 Presentation on LipNet: End-to-End Sentence-level Lipreading
Presented by Sanullah Manzoor PhdCS-16003
Auto-Encoder
Apr 20, 2017 Slides
Hugo Larochelle’s lectures are treat
Generative Adversarial Networks
Apr 25, 2017 Generative Adversarial Networks Paper
An introduction to Generative Adversarial Networks (with code in TensorFlow) Introductory guide to Generative Adversarial Networks (GANs) and their promise!
Video Analysis using Deep Learning
Apr 27, 2017 Understanding different architectures used for the Video Analysis

Papers Presented by Students

  1. Accurate Image Super-Resolution Using Very Deep Convolutional Networks – CVPR 2016
  2. DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations – CVPR 2016
  3. End to End Learning for Self-Driving Cars – CVPR 2016
  4. Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
  5. Domain Adaptation for Large-Scale Sentiment Classification: A Deep Learning Approach – ICML 2011
  6. Lip Reading Sentences in the Wild – CVPR 2017
  7. Harnessing Object and Scene Semantics for Large-Scale Video Understanding – CVPR 2016
  8. Mask R-CNN – CVPR 2016
  9. Short Text Clustering via Convolutional Neural Networks – NAACL-HLT. 2015
  10. Recurrent Convolutional Neural Networks for Text Classification – AAAI 2015
  11. Neural responding machine for short-text conversation – arXiv 2015
  12. LipNet: Sentence-level Lipreading – arXiv 2016
  13. A persona-based neural conversation model – arXiv
  14. Sequence to sequence learning with neural networks – NIPS 2014
  15. A machine learning approach to visual perception of forest trails for mobile robots – IEEE Robotics and Automation Letters 1.2 2016