Structure

At a high-level:

• Form groups of 2 people people.
• Pick a paper from this list (access to UC Davis emails only).
• Formulate insightful and impactful follow-up questions to the paper.
• Make initial strides in solving these questions, via theory or experiments.

Details

The project constitutes 30% of your final grade. Your grade on the project will be evaluated by the following rubric.

Project milestone 1 (5%) - due January 31

Send an email to Spencer (sfrei@ucdavis.edu) with the subject "STA 250 Final Project", and all group members on CC, with the following info:

• Name of the paper you are considering.
  • The paper must be from this list (you need to access using your UC Davis account). Papers which have already been taken by another group will be marked on the spreadsheet, you must pick one which has not yet been taken.
  • If you choose an empirical paper, your project will likely involve a significant experimental component, so be sure that the group members are very comfortable with (or willing to get up to speed with) using PyTorch/JAX/etc.
• Names of group members.
• A summary of the results in the paper in 2-3 paragraphs.

Project milestone 2 (15%) - due February 12

Prepare a PDF with the following info:

• Name of the paper you chose.
• Names of group members.
• A summary of the results in the paper in 2-3 paragraphs.
• 2-3 interesting and impactful follow-up questions to explore. Potentially relevant things to consider to help formulate these questions:
  • Are there any assumptions which seem unusual, unjustified, or too stringent? What happens if you remove these assumptions, does the proof break down? If you run some experiments under different assumptions, do they still hold?
  • Experiments they provide do not seem to match what the theory suggests, or are not reproducible.
  • They are attempting to solve the "wrong" problem, when there is a very related/more relevant problem whose solution would be more revealing and interesting.
  • You think of a tweak to their setup, run some experiments, and find a new/surprising/contradictory behavior which is worth investigating further.
  • Can the proof be generalized to a more complex and interesting setting? What fundamentally breaks down if you try to extend it (e.g. from linear classifiers to neural nets, smooth to non-smooth activation functions, noiseless labels to noisy labels, training two layers vs training first layer only, etc.)
• A plan for how you will approach at least one of these problems, and the division of labor between the two group members.

The proposal should be submitted as a PDF on Canvas. The proposal should be at most 2 pages long. No late submissions will be accepted. It is highly recommended that you meet with Spencer after class or during office hours to discuss before submitting the proposal.

Project presentation (20%)

You will present your presentation using slides (see here for some example talks). The presentation will be 20 minutes long: 15 minutes of presentation, 5 minutes for questions. A good rule of thumb is that you should spend at least 1, ideally 2 minutes per slide with math on it. It is better to simplify some of the math to provide cleaner slides with fewer equations than to try to present everything in full detail on the slides.

Your grade will be based on:

• Clarity of slides
• How well key messages are delivered within the alotted time
• How well you handle audience questions

I use Beamer to make slides in LaTeX, others use Powerpoint or Google Slides or Keynote. Make sure that for the day of your presentation, you have a PDF version of your slides which you can use.

Attendance to project presentations (10%)

Attendance at all project presentations is mandatory. For each missed presentation day, you will lose 5 percentage points.

Project report (50%)

Each team will submit a written project report in PDF, via LaTeX using the NeurIPS 2023 style files. The report should be at most 9 pages in the main section, with an unlimited number of pages for references and the appendix. A concise/succinct report which clearly communicates your work and findings is better than a long-winded one.

The report should be in the style of a standard ML conference paper (see, e.g., this or this): an introduction, literature review, background/preliminaries/motivations, main findings, and discussion/conclusion.

The report will be graded equally on:

• Clarity (well-written, well-organized, clear description of motivations, methodology and theory well-described)
• Literature survey (is related work adequately discussed)
• Correctness
• Novelty, substance, and findings (were there significant scientific insights?)

Each member of the team will receive the same grade for the report.