Assignments and Projects
Programming Assignment 1 — Scalable Transformer Training & Profiling
The goal of this assignment is to give you hands-on experience with training a modern Transformer-based model, profiling its computational performance, and applying systems-level optimizations to improve efficiency.
Programming Assignment 2 — Policy Gradients with Intrinsic Exploration
Implement PPO/A2C with intrinsic motivation (e.g., RND/ICM) for a sparse-reward game; log extrinsic/intrinsic rewards and compare against a no-intrinsic baseline.
Programming Assignment 3 — Neuroevolution in MuJoCo
Evolve policies with GA or ES on continuous-control tasks; optionally add novelty/QD and compare convergence/diversity/sample-efficiency to PPO.
Paper Presentation and Review
A requirement of Foundations of Machine Learning for Systems Researchers is for each student to present a research paper to the class. This assignment involves conducting a careful review of the paper and leading an in-class discussion of its contributions and relevance. Each student will be responsible for presenting one paper and moderating the associated discussion. Presentations will be 30 minutes in length, followed by a 5-minute review delivered by another student. Presentation slides must be submitted to Canvas two days prior to the scheduled presentation date (by end of day) to allow for feedback and potential revision.
Students assigned as reviewers are expected to prepare a short critique of the presented paper in the style of a NeurIPS review (2025 guidelines can be found here for reference). An example review for the FlashAttention paper can be found here.
Final Project
Group project to address a problem in systems using paradigms introduced in the course, such as deep learning, reinforcement learning, evolutionary algorithms, and other paradigms. Novelty, evaluation quality, and clarity matter most.
- Proposal + related work; interim check-ins; final paper (conference style, ICLR preferred) and in-class presentation.
See the Schedule for official deadlines.