← Teaching

Special Topics: Machine Learning for Scientists

PHYS-GA 2043

Goals of the class

This course has four goals. First, that you experience some wonder: how can these methods possibly work? And that you turn this wonder into understanding, by deriving and implementing the core methods yourselves rather than importing them. Second, that you begin to think like a machine learner, not just a user of tools: to see modeling choices, inductive biases, and failure modes as things you design. Third, that you can read current ML research papers with confidence and healthy skepticism. And finally, that you can develop interesting projects of your own at the intersection of machine learning and your field, because that intersection is where the best ideas in both directions now come from.

Schedule

Tue/Thu 4:00–5:15 pm · Fall 2026 · Dates subject to change.

Assignments

Labs

Five labs, done either individually or in pairs, each implementing a core method from scratch. Each lab is due the day of its in-class peer review (see the schedule above).

• Lab 1: MLPs from scratch (numpy/jax). Build an MLP and implement backprop yourself. Teaches what training is.
• Lab 2: Transformers from scratch (numpy/jax). Implement attention, build a small transformer.
• Lab 3: Flow matching from scratch (torch/jax). Regress a vector field on 2D toy data, sample by integrating an ODE, evaluate exact likelihoods.
• Lab 4: Simulation-based inference (torch/jax). Condition your Lab 3 flow on observations: neural posterior estimation. Validate your model with predictive checks, calibration, and coverage, including a deliberately broken setup you must catch from the diagnostics alone.
• Lab 5: Reinforcement learning (torch/jax). Implement REINFORCE, land a lunar lander.

Peer review

Each lab is reviewed in class by another pair; submit before class. Answer:

• What did they do that's interesting or clever?
• Where did they choose differently than you, and why do you think?
• What's one thing you don't understand about their code?
• Can you find a way to break it?

Project

The goal of the course project is to produce a piece of original research at the intersection of machine learning and physics, the kind of work you could submit to a NeurIPS workshop. The final deliverable is a 4-page paper in NeurIPS workshop format (references excluded), presented at the end of the semester. You'll get there through four checkpoints:

• Literature review (due Thu Sep 17). Curate a list of papers in your area of interest: what they did, what surprised you, and what you would ask next. Form your team.
• Research proposal (due Sun Oct 4). At most two pages (shorter is fine): your question in one sentence, why your subfield cares, the data or simulator you can actually touch, the method, and a plan for the remaining weeks. Proposals are peer-reviewed by another team the following week, and you pitch the revised plan to the class on Thu Oct 15.
• Money plot check (due Fri Oct 30). A short mid-project report: your question in one sentence (has it changed?), a sketch of your money plot, the version you hope for and the version you fear (drawn is fine; the concept is what matters), one real plot from actual data, however ugly, your main blocker, and your plan B.
• Final presentations & report (exam slot, Dec 16–22). Present the project, with the October sketch next to the final result, and hand in the 4-page paper the same day.

Resources for the literature review

A personal, non-exhaustive starting point, heavily biased toward astrophysics and cosmology. It is meant to be inspirational.

To find work in your own area, browse past editions of two workshops:

• Machine Learning and the Physical Sciences (NeurIPS): 2025 · 2024 · 2023 · 2022 · 2021 · all editions
• AI for Science (NeurIPS & ICML): all editions