About the Course and Prerequisites
This course will study the sources and measures of unfairness stemming from machine learning, as well as possible interventions to alleviate said unfairness. The class will contain a mix of assignments: there will be 3 assignments. Each assignment will contain 3 parts: a programming task, several mathematical questions which will require rigorous proofs to answer, and an analysis/writing component.
Tentative list of topics (a (!) indicates that the topic may not be covered):
Statistical vs Individual definitions of fairness
Sources of Unfairness in ML
Tensions between different fairness notions
Fairness interventions and long-term effects of those interventions
Equilibria: does competition “drive out” any unfairness a business’s model might have?
Unfair data vs unfair objectives or algorithms
Thinking outside the (CS) box: what do other communities think unfairness means, and how have they grappled with these questions?
Prerequisites: The course will assume a great deal of comfort with introductory machine learning. It will also assume familiarity with concentration results such as Chernoff/Hoeffding.
There will not be a textbook for the course.
Our discussion will be guided by papers, monographs, and lecture notes that are available online.
An incomplete list that will grow:
Discussion Forum and Email Communication
There will be a Slack channel (first day of class). This is your first resource for questions. For private or confidential questions email the instructor.
Grading and Evaluation
(30%) 3 homeworks
each will have 1 programming task, 2-3 problems requiring proofs, and 1 paper to read/write a critique, refutation, or supporting argument for.
(30%) In-class discussion
Each class will have readings, there will be several questions during lecture related to the readings. Contribute to some of these discussions
(40%) Final Project
This project can be empirical (awesome!) or formal or some combination thereof. Can be done solo or in pairs.
Scribe notes should be prepared using the Latex template. The final notes should be turned in within a few days following lecture with the understanding that the majority of the notes should be completed before lecture.
- Jan. 6
- Welcome, logistics, overview of course topics
- Jan. 8
- What assumptions do we make about our world, our data, our goals?
- Jan. 13
- Legalese and (some very limited) historical context.
- Jan. 15
- Jan. 20
- Jan. 22
- Statistical Notions of Fairness
- Jan. 27
- Virtual Class, part 1.
- Reading: Please read 1 paper appearing at fatconference.org for 2019 or 2020, and outline the primary perspective/strengths/weaknesses of said approach. Please post this to slack by midnight on Tuesday Jan 28. Look for the thread the instructor will set up to post your summary.
- Jan 29
- Virtual Class, part 2.
- Please comment on at least 2 of your classmates' summaries from Tuesday, with either questions, additional related research, or another discussion point regarding the research direction.
- Feb 3
- Real-valued predictions, calibration, and ROC curves. Homework 1 released
- Reading: No addtional reading.
- Feb 5
- Interpolating between Individual and Group Fairness Measures.
- Feb 10
- Long-term outcomes of short-term fairness constraints.
- Feb 12
Introduction to learning from experts and bandit learning, Feedback loops
- Feb 17
- No Class, President's Day
- Feb 19
- Causality and Fairness 1.
- Causality and Fairness 2.