Deep Unsupervised Learning

Spring 2019

Note: Spring 2020 offering of the course is hosted here

About: This course will cover two areas of deep learning in which labeled data is not required: Deep Generative Models and Self-supervised Learning. Recent advances in generative models have made it possible to realistically model high-dimensional raw data such as natural images, audio waveforms and text corpora. Strides in self-supervised learning have started to close the gap between supervised representation learning and unsupervised representation learning in terms of fine-tuning to unseen tasks. This course will cover the theoretical foundations of these topics as well as their newly enabled applications.

Instructors: Pieter Abbeel, Peter Chen, Jonathan Ho, Aravind Srinivas

Communication: https://piazza.com/berkeley/spring2020/cs294158

Communication: https://piazza.com/class/jr8swguu59nem

(Alternative: cs294-158-staff@lists.berkeley.edu)


When: Wednesdays 5-8pm; first lecture on 1/30

Where: Moffitt Library 145 [CalID required to entire library]

Tentative list of topics (much subject to change):

Generative adversarial networks, variational autoencoders, autoregressive models, flow models, energy based models, compression, self-supervised learning, semi-supervised learning.

Prerequisites: significant experience with probability, optimization, deep learning

Office Hours

(starting week of 2/4)

Jonathan: Mondays 4-5pm 734 Sutardja Dai

Aravind: Tuesdays 9-10am 734 Sutardja Dai

Pieter: Thursdays 9-10am 746 Sutardja Dai [but there will be no office hours on Thu 4/11 and Thu 4/18; instead there will be office hours Wed 4/17 11-noon]

Peter: Thursdays 6-7pm 734 Sutardja Dai


HW1: Autoregressive Models (due 2/11) HW1.PDF, HW1_template.tex, mnist-hw1.pkl, distribution.npy

HW2: Flow Models (due 2/26) HW2.PDF, HW2_template.tex, hw2_q2.pkl

HW3: Latent Variable Models (due 3/14) HW3.PDF, HW3_template.tex, hw3_q2.pkl

HW4: Implicit Models (due 4/9) HW4.PDF, HW4_template.tex

Tentative Schedule / Syllabus

Week 1 (1/30) [youtube]

Lecture 1a: Logistics

Lecture 1b: Motivation

Lecture 1c: Likelihood-based Models Part I: Autoregressive Models

Week 2 (2/6) [youtube]

Lecture 2a: Likelihood-based Models Part I: Autoregressive Models (ctd) (same slides as week 1)

Lecture 2b: Lossless Compression

Lecture 2c: Likelihood-based Models Part II: Flow Models

Week 3 (2/13) [youtube]

Lecture 3a: Likelihood-based Models Part II: Flow Models (ctd) (same slides as week 2)

Lecture 3b: Latent Variable Models - part 1

Week 4 (2/20) [youtube]

Lecture 4a: Latent Variable Models - part 2

Lecture 4b: Bits-Back Coding

Week 5 (2/27) [youtube]

Lecture 5a: Latent Variable Models - wrap-up (same slides as Latent Variable Models - part 2)

Lecture 5b: ANS coding (same slides as bits-back coding)

Lecture 5c: Implicit Models / Generative Adversarial Networks

Week X (3/6)

Final Project Discussion

Week 6 (3/13) [youtube]

Lecture 6a: Implicit Models / Generative Adversarial Networks (ctd) (same slides as 5c)

Lecture 6b: Non-Generative Representation Learning [UPDATED 3/24]

Week 7 (3/20) [youtube]

Lecture 7: Non-Generative Representation Learning (same slides as 6b)

Spring Break Week (3/27)

you are on your own :)

Week 8 (4/3) [a,b: youtube; c: youtube]

Lecture 8a: Strengths/Weaknesses of Unsupervised Learning Methods Covered Thus Far

Lecture 8b: Semi-Supervised Learning

Lecture 8c: Guest Lecture by Ilya Sutskever

Week 9 (4/10) [a: youtube; b: youtube]

Lecture 9a: Unsupervised Distribution Alignment

Lecture 9b: Guest Lecture by Alyosha Efros

Week 10 (4/17) [youtube]

Lecture 10: Language Models (Alec Radford)

Week 11 (4/24) [youtube]

Lecture 11: Representation Learning in Reinforcement Learning

Week 12 (5/1) [youtube]

Lecture 12: Guest Lecture by Aaron van den Oord [slides not available]

Week 13 (5/8)

RRR week: no lecture

Week 14 (5/15)

Final Project Presentations


Q: How do I get into this course?

A: Please fill out this survey, which we will use for admissions. (deadline passed)

Q: Can undergraduates take this course?

A: This course is targeted towards a PhD level audience. But certainly exceptional undergraduates could be good fits, too, and your ability to take this course is not directly affected by your grad/undergrad student status, but by things we measure in the survey and by your performance on Homework 1.

Q: Is this a real course or a seminar?

A: This is a real course. Instructors will give most of the lectures. There will be substantial homework. There will be a substantial final project.

Q: How will grading work?

A: Details to be determined. But we expect grades to be determined largely by 3-5 substantial homeworks + a final project.

Q: I already want to start learning now, what can I do, can you point me to some research papers maybe?

A: Certainly, here is a zip file with about 100 papers much on-topic for this course, happy readings!!