Cornell CS 6784 Fall 2023
Meeting MW 2:55-4:10pm in 255 Olin Hall
Professor Sarah Dean, informal office hours MW after lecture in 255 Olin or MW 4:30-5pm in Gates 416A by appointment
TA Raunak Kumar, Office hours Thursdays 2:30pm - 3:30pm in Rhodes 576
Please use Ed Discussions for any questions.
Please fill out this form to join the class's collaborative GitHub repository.
If you are on the waitlist, please fill out this interest form.
Feedback between machine learning models and the environment in which they are deployed leads to a host of challenges, from distribution shift to bias to polarization. This graduate level course will introduce theoretical foundations for studying such phenomena. We will cover the frameworks of online/adaptive learning, control theory, and reinforcement learning. For each, we will discuss algorithms for ensuring properties like stability, robustness, safety, and fairness. We will also discuss the social and ethical concerns which motivate these algorithms and properties. Paper discussions and a research project are major parts of the course.
Unit 1: Learning to Predict (Aug-Sept)
Topics: Supervised Learning (Fairness), Online Learning, Dynamical Systems (Stability)
Unit 2: Learning to Act (Sept-Oct)
Topics: Multi-Armed Bandits, Optimal Control & Reinforcement Learning (Robustness), Model Predictive Control (Safety)
Unit 3: Student-led Paper Discussions (Oct-Dec)
The detailed calendar will be updated throughout the semester. List of references and papers will be posted in September.
Knowledge of ML at the level of CS4780 is recommended. Perhaps more important is mathematical maturity and a working understanding of linear algebra, convex optimization, and probability. The following references may be useful to review: Linear Algebra Review and Reference, Convex Optimization Overview, and Review of Probability Theory.
Students will complete weekly assignments, present a paper, and work on a project during the semester. Depending on enrollment, some of this work may be done in pairs or groups.
Assignments will be posted each week on Wednesday and are due the following Wednesday. We will use GitHub collaboration tools to manage and collect your work.
During the second half of the semester, students will present selected papers (list to come) and lead a discussion. Students are required to schedule a meeting with the TA to go over their presentation at least two days before they are scheduled to present. Presentation Details.
Projects can be done in groups of up to four, with expectations scaling with the size of the group. Students are encouraged to propose a topic that connects class material to their research. The deliverables are:
- Project proposal (1 page) due October 6
- Midterm update (3 pages) due November 10
- Project report (5-6 pages) due last day of class
Students will be evaluated by:
- 50% final project
- 20% paper presentation
- 20% weekly assignments
- 10% participation