General Informaion

• Instructor:  Michael P. Kim
• Meetings:  Mon/Wed, 10:10-11:25a
• Location:  Hollister 362
• TAs:  Princewill Okoroafor and Mohit Gurumukhani
• Office Hours:
  • Michael: Wed, 2p (or by appointment) in Gates 320
  • TAs: Wed, 3p in Theory Lab (Gates 324 / 336)

Course Description

Predictive algorithms influence and shape society. The use of machine learning to make predictions about people raises a host of basic questions: What does it mean for a predictive algorithm to be fair to individuals from marginalized groups? On what basis should we deem a predictive algorithm to be valid? And when should we trust (or distrust) a predictor’s output? This course surveys recent developments in the theory of responsible machine learning. We overview new paradigms for formulating learning problems and highlight key algorithmic tools in the study of fairness, validity, and robustness.

Topics covered include:

• Multicalibration
• Outcome Indistinguishability
• Omniprediction
• Performative Prediction
• Distributional Robustness
• Verification of Learning

Objectives and Prerequisites

The course is intended for early-career PhD students, who are pursuing research in CS Theory, Machine Learning, and related fields. After completing the course, students should be able to:

• Identify common patterns and assumptions underlying modern prediction problems;
• Evaluate, given new settings, whether using machine prediction is appropriate;
• When appropriate, apply principled frameworks to reason about machine learning responsibly.

Students are expected to have prerequisite familiarity with machine learning (at the level of CS 3780) and theory of computation (at the level of CS 4820). The course approaches topics rigorously, so mathematical maturity is required.

Evaluation

Students will be evaluated based on the following criteria.

• Attendance and participation throughout the semester
• periodic homework assignments
• a comprehensive research project