Course Information

About

As an exciting new field, deep unsupervised learning has gradually emerged as a promising alternative to supervised approaches to representation learning -- with some practical and theoretical considerations. To begin with, unsupervised data is much cheaper to obtain, but more importantly, as humans, we don't need millions of labeled data to learn.

This class will provide an in-depth and comprehensive overview of the fundamental concepts and recent advances in the field of deep unsupervised learning. The first part of the course focuses on deep generative models such as autoregressive models, normalizing flow models, variational autoencoders, generative adversarial networks and their extensions with discrete latent variables. The second part covers self-supervised learning, including pretraining of large language models. The class is mostly modeled after the Deep Unsupervised Learning course at Berkeley. The class is taught by Aykut Erdem.

Instruction style: During the semester, students are responsible for studying and keeping up with the course material outside of class time. These may involve reading particular book chapters, papers or blogs and watching some video lectures.

Time and Location

Lectures: Monday and Wednesday at 08:30-09:45 (online via Zoom)

Communication

The course webpage will be updated regularly throughout the semester with lecture notes, presentations, assignments and important deadlines. All other course related communications will be carried out through Blackboard.

Prerequisites

COMP547 is intended for graduate students enrolled in Computer Science MS and PhD programs. Senior undergraduate students and all non-COMP graduate students need the instructor's permission to register for the class. The prerequisites for this course are:

  • Programming (you should be a proficient programmer to work out the assignments and to implement your course project.)
  • Calculus (differentiation, chain rule) and Linear Algebra (vectors, matrices, eigenvalues/vectors) (MATH107)
  • Basic Probability and Statistics (random variables, expectations, multivariate Gaussians, Bayes rule, conditional probabilities) (ENGR200)
  • Machine Learning or Deep Learning (you can still survive this course without a machine learning course before, but it is highly recommended (ENGR421, COMP541)
  • Optimization (cost functions, taking gradients, regularization)

Course Requirements and Grading

Grading will be based on

  • Assignments (30%) (3 assignments x 10% each)
  • Midterm Exam (10%)
  • Course Project (presentations and reports) (37.5%),
  • Paper Presentations (18%),
  • Paper Reviews (4.5%),

Reference Books

  • Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning, MIT Press, 2016 (draft available online)

Schedule

Week Topic Assignments
Feb 15-17 Introduction to the course (Survey)
Neural Building Blocks I: Spatial Processing with CNNs
Feb 22-24 Neural Building Blocks II: Sequential Processing with RNNs
Neural Building Blocks III: Attention and Transformers
Mar 1-3 Autoregressive Models Assg 1 out
Mar 8-10 Normalizing Flow Models
Mar 15-17 Variational Autoencoders Assg 1 due, Assg 2 out
Mar 22-24 Generative Adversarial Networks
Mar 29-31 Generative Adversarial Networks (cont'd) Assg 2 due, Assg 3 out
Apr 5-7 No classes - Spring Break
Apr 12-14 Discrete Latent Variable Models Project proposal due
Apr 19-21 Strengths and Weaknesses of Current Models Assg 3 due
Apr 26-28 Self-Supervised Learning
May 3-5 Pretraining Language Models
May 10-12 Project Progress Presentations Project progress reports due
May 17 Project Progress Presentations
May 22 (11:45am) Midterm Exam (guide)
May 24-26 Self-Supervised Learning for Vision and Language
Jun 7-9 Final Project Presentations Final project reports due
Detailed Syllabus and Lectures