You can propose any project that involves vision transformers in any area of interest. Projects can be completed individually or in groups of 2-3. For GPU access, please create an account on Longleaf. Some examples of what a good project might look like:
Propose a modification to an existing method and show whether it helps.
Build a benchmark for a capability that current models are not evaluated on.
Take an idea from one area of the course and apply it where it has not been tried.
Run a systematic comparison that no paper has done.
Build a system that does something no existing model does.
You are not limited to these examples. The best path toward a successful project is to be creative and work on something that you feel excited about. You are encouraged to use coding agents and other AI tools. Include a brief note on which tools you used and how you used them with each submission (see Course Policies). Projects are evaluated on the final result and what you learned from building it. A well-executed negative result with a clear explanation is acceptable.
If you are unsure whether your idea is suitable, come talk to me.
The project is worth 40% of your final grade.
Your initial proposal presentation should cover the following items:
The problem that you are trying to solve.
The motivation behind the problem.
Your proposed approach and how it relates to prior work.
The experiments that you plan to conduct.
The datasets that you plan to use.
You project milestone presentation should extend your project proposal to include the following:
A preliminary set of results.
Your analysis of those results.
Additional experiments that you plan to run.
Along with your presentation, submit a link to your code repository.
Your final project presentation should include the following:
The final set of results.
Your analysis of those results.
Your overall conclusions and findings from the project (what did you learn?)
Along with your presentation, submit a link to your code repository. This is the main deliverable.