Technological advances in how information is generated, collected, managed, and analyzed raise a variety of societal concerns. This course will explore how to use mathematical methods to articulate some of these challenges formally, reason about them rigorously, and design algorithms to mitigate them. Potential topics include data privacy, fairness in algorithmic learning and decision-making, evaluation and interpretation of complex machine learning models, feedback loops in data-driven systems, and strategic behavior. An emphasis will be placed on understanding the challenges arising from modern machine learning, such as through large language models.
The course, which is aimed at advanced undergraduates in computer science, data science, ECE, or statistics, will engage with both technical components of the problem area (involving programming and theoretical problems) as well as questions in policy and ethics (involving reading, discussing, and writing about papers from those areas).
| Instructor: | Mark Bun, mbun [at] bu [dot] edu | |
| Instr. Office Hours: | Tue 3:30-4:30 PM (CDS 1021) | |
| Wed 4-6 PM (CDS 1021) | ||
| Teaching Fellow: | Mandar Juvekar, mandarj [at] bu [dot] edu | |
| Mandar's Office Hours: | Mon 9-10 AM (CDS 10th Floor Yellow Lounge) | |
| Thu 12:45-1:45 PM (CDS 10th Floor Yellow Lounge) | ||
| Class Times: | Tue, Thu 2:00-3:15 PM (PSY 212) | |
| Discussion Sections: | Mon 12:20-1:10 PM (CDS 701) | |
| Mon 1:25-2:15 PM (CDS 701) |
Course Website: https://cs-people.bu.edu/mbun/courses/391_F25. The website contains the course syllabus, schedule with assigned readings, homework assignments, and other course materials.
Piazza: https://piazza.com/bu/fall2025/cs391b. All class announcements will be made through Piazza, so please set your notifications appropriately. You can also find course handouts and homework/test solutions there. Please post questions about the course material to Piazza instead of emailing the course staff directly. It is likely that other students will have the same questions as you and may be able to provide answers in a more timely fashion. Active participation on Piazza may add extra points to your participation grade.
Gradescope: https://gradescope.com. Sign up for a student account on Gradescope using your BU email address. The entry code for the course is 6KYJPK. Homework assignments are to be submitted to Gradescope in PDF format.
Probability (e.g., CS 237 or DS 122), algorithms (e.g., CS 330 or DS 320), and the basics of machine learning (e.g., CS 365 or DS 340). Additional background in AI, machine learning, or statistics is helpful but not required. Students from programs outside CS and DS are welcome but should contact the instructors regarding prerequisites.
| Date | Topics | Reading/Reference | Handouts/Assignments |
|---|---|---|---|
| Tue 9/2 | Welcome, introduction, supervised learning basics | [BHN] pp.44-51, [SS] Ch. 2 | |
| Thu 9/4 | Sources of bias in data and models, philosophical foundations of fairness, start group fairness | [BHN] Ch. 1, pp.52-63, 76-80; Machine Bias + analysis Optional: Fairness in Philosophy |
|
| Tue 9/9 | Group fairness and how to achieve it | [BHN] pp.52-63, 67-71, 81-87, 92-94, 97-101 Equality of Opportunity in Supervised Learning |
|
| Thu 9/11 | Group fairness tradeoffs | [BHN] pp.63-67 Optional: What Should We Do When Our Ideas of Fairness Conflict? |
|
| Tue 9/16 | Individual fairness | Fairness through Awareness, [BHN] pp.94-97 | |
| Thu 9/18 | Multi-group fairness | Calibration for the (Computationally-Identifiable) Masses | |
| Tue 9/23 | Fairness and causality | [BHN] pp.104-121 | |
| Thu 9/25 | Counterfactual fairness | [BHN] pp. 121-131 Counterfactual Fairness |
HW 1 due |
| Tue 9/30 | Intro to interpretability | [Mol] Ch 1-4 Towards a Rigorous Science of Interpretable Machine Learning |
|
| Thu 10/2 | AI in government, linear and logistic models | [Mol] Ch 6-7 | |
| Tue 10/7 | Decision rules, sets, lists, trees | [Mol] Ch 9-10 | |
| Thu 10/9 | Local model-agnostic explanations | [Mol] Ch 12-14, 17-18 | |
| Tue 10/14 | NO CLASS — Substitute Monday for Indigenous People's Day | ||
| Thu 10/16 | Counterfactual explanations, global model-agnostic explanations | [Mol] Ch 15, 19-24 | HW 2 due |
| Tue 10/21 | Test Review | ||
| Thu 10/23 | Test 1 | ||
| Tue 10/28 | Interpreting convolutional networks | [Mol] Ch 27-28 Zoom In: An Introduction to Circuits |
|
| Thu 10/30 | Interpreting transformers | A Mathematical Framework for Transformer Circuits In-context Learning and Induction Heads |
|
| Tue 11/4 | Induction heads, locating facts in LLMs | In-context Learning and Induction Heads ROME: Locating and Editing Factual Associations in GPT |
|
| Thu 11/6 | Activation steering | Steering Language Models with Activation Engineering | |
| Tue 11/11 | Emergent misalignment | Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs | |
| Thu 11/13 | Alignment via reinforcement learning | Training Language Models to Follow Instructions with Human Feedback | |
| Tue 11/18 | Student presentations (HW 3) | ||
| Thu 11/20 | Student presentations (HW 3) | ||
| Tue 11/25 | Attacks and memorization | Exposed! A Survey of Attacks on Private Data Extracting Training Data from Large Language Models |
|
| Thu 11/27 | NO CLASS — Thanksgiving | ||
| Tue 12/2 | Differential privacy basics | [DR] Ch 1, 2, 3.1-3.3, 3.5 | |
| Thu 12/4 | Differentially private algorithms | [DR] Ch 3.3-3.5 | |
| Tue 12/9 | Private model training | Differentially Private Empirical Risk Minimization | HW 4 due |
| Tue 12/16 | Test 2 (3-5PM), PSY 212 |
You do not need to buy a textbook for this class. Required reading will come from sources that are either freely available on the web or accessible using your BU institutional login. The abbreviated references in the schedule are to the following textbooks and monographs.
[BHN] Solon Barocas, Moritz Hardt, and Arvind Narayanan. Fairness and Machine Learning: Limitations and Opportunities. [link] [DR] Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential Privacy. [link] [FV] Ferdinando Fioretto and Pascal Van Hentenryck (ed.). Differential Privacy in Artificial Intelligence: From Theory to Practice. [link] [Mol] Christoph Molnar. Interpretable Machine Learning: A Guide to Making Black Box Models Explainable. [link] [MRT] Mehryar Mohri, Afhin Rostamizadeh, and Ameet Talwalkar. Foundations of Machine Learning. [PDF and e-book] [SS] Shai Shalev-Shwartz and Shai Ben-David. Understanding Machine Learning: From Theory to Algorithms. [PDF]There will be roughly four homework assignments, tentatively due 9/25, 10/16, 11/13, and 12/9, each at 11:59PM. Assignments will include a mix of mathematical problem-solving, programming, and sociotechnical writing. Some homework problems will be designed to be challenging or open-ended. You will want to start early to give yourself time to think deeply about the problems. Even so, do not be discouraged if you cannot completely solve all of the problems, and we welcome well-reasoned partial ideas toward solutions. We will aim to turn around homework solutions shortly after each due date, so we cannot accept late assignments without prior arrangement due to exceptional circumstances.
You are allowed, and indeed encouraged, to collaborate with other students on solving most of the homework problems. However, you must write the solutions independently in your own words. Details of the collaboration policy may be found below under the header "Collaboration Policy."
You may want to use LaTeX to typeset your homework solutions. LaTeX is the standard document preparation system used in the mathematical sciences. Using LaTeX makes it easier for you to revise and edit your solutions and for us to read them, so you will never lose points for illegibility.
My preferred LaTeX editors are TexShop for Mac and TexStudio for Windows. If you would like to give LaTeX a try on the web without installing anything on your computer, Overleaf is a good option.
Not so short intro to LaTeX. A LaTeX tutorial.
Two in-class tests are scheduled for Thursday, October 23 and (tentatively) Tuesday, December 16 during our final exam slot. (Please wait until the official University final exam schedule is finalized before making your end-of-semester travel plans.) Each test will cover roughly half of the course content. Test 2 is not intended to be cumulative.
You may bring one double-sided 8.5" x 11" sheet of notes to each test. Note sheets may be either handwritten or typeset. You may not use any other aids during the exam, including but not limited to books, lecture notes, calculators, phones, or laptops.
Your active participation in class and in discussion section activities is an essential part of your learning and will make it much more enjoyable for everyone. You can also earn participation credit by asking thoughtful questions on Piazza or during office hours. Midway through the semester, I will send you an indication on how your participation in class is going.
You may verbally collaborate on homework problems with up to 3 students currently enrolled in the course. However, you must write your solutions independently in your own words, and understand what you are writing well enough to explain your solutions to the course staff if asked. You must also list the names of everyone with whom you collaborated on your assignment.
You may use external resources such as textbooks, lecture notes, research articles, blog posts, and videos to supplement your general understanding of the course topics or conduct research as needed. You must cite your sources for any ideas or auxiliary materials (e.g., datasets) you submit that are not your own. You may also get help from anyone on debugging LaTeX or other issues that are clearly more general than the scope of the course.
You may not look up answers to homework problems in the published literature, on the web, or using generative AI. You may not submit work that you did not write yourself. You may not share written work with anyone else.
The use of generative AI tools like ChatGPT is permitted as long as it conforms to the policies above. That is, you may use them for general research (as you might use a search engine), but you may not prompt them to answer specific homework problems and you may not misrepresent their output as if it were your own work.
If you are uncertain as to whether a particular kind of interaction with someone else constitutes illicit collaboration or academic dishonesty, please ask the instructor before taking any action that might violate the rules. If you can’t reach the instructor in time, cite your sources and include a clear explanation of what happened on your submission.