Project Logistics

• Use an LLM for everything — coding, execution, debugging, plots, math — except the report.
• Proposal (1 page max) due Oct 13.
• Checkpoint (1 page progress report) due Nov 13.
• Report due Dec 6.
• Report is written entirely on your own, no LLM. You must be able to explain it.
• No teams — individual projects only.
• Scope must exceed a homework problem — this cannot be a simple exercise at HW difficulty. Expected total effort: about 20-30 hours, plus writing the report.
• Instructor/TAs can suggest ideas for Tracks 1 and 2, but their suggestions are no better than what you can find yourself.
• Report must include a section comparing your result against an external ground truth (paper's numbers, a baseline, a bound, a hand-checked case), if feasible for your project.
• Grading: simple 0 / half / full schema (see below).

4 Tracks (pick one)

1. Replicate and critique a recent paper

Pick a paper with a significant ML idea/method from roughly the last 4 years. Reproduce its central result, at whatever scale is feasible. Identify and compare against the 3-5 most relevant related works on that same problem/task/setup — not just whatever the target paper itself happens to cite. Explain why the new idea helps conceptually. If your numbers don't match the paper's, explain why.

2. Theory vs. practice

Pick a theoretical claim from class or the literature — a generalization/error bound, a convergence rate, a stopping/termination criterion, a closed-form optimum. Implement and run the relevant algorithm, then measure how well the claim actually holds: is the bound tight or loose, does the observed rate match the predicted one, does the closed form match the trained/converged values. Report and explain the gap, if any.

3. Build a pipeline for a real problem

Pick a real problem you care about, with real data — prefer a dataset that's already available over one you'd need to scrape or clean from scratch, which tends to eat the whole time budget by itself. Build a pipeline combining algorithms from class. Be ready to defend your design choices: why this model, why this metric, what would break at 10x the data, what the error analysis actually showed.

4. Propose your own

If you already have a suitable project or research interest, propose it. Must match the scope of the other tracks and be approved by the instructor in the proposal.

Proposal

The proposal's job is to prove the project is actually doable in the time you have — not to describe the result you expect to get.

• If data is involved, have it downloaded already, or demonstrate that you have access to it.
• Track 1: show that everything the paper depends on — implementation details, proprietary data or standards, specific hardware — is either publicly available or something you can obtain or work around. If an ingredient is missing or locked down, say so and how you'll handle it.
• Track 3: show that data access, compute/hardware, and LLM usage/credits are all available, or that you have a concrete plan for resolving whichever one isn't yet.
• Track 4: satisfy whichever of the above feasibility checks applies to what you're proposing.

Grading

Project is about 15% of the final grade.

• Full: specific, non-generic project; a real comparison against ground truth (paper's numbers, a baseline, a bound, a hand-checked case); a report that shows you understand the whole thing, in your own words.
• Half: project ran and produced results, but the comparison-to-ground-truth is thin/missing, or the report doesn't reflect real understanding.
• Zero: generic project, no real comparison, or a report that is (or reads as) LLM-written.