The Final Project for CS217 will be an open-ended assignment done in teams of at most 4, with focus on things covered in class. For your project, you may choose any of the following.
- Implement an existing model such as ResNet, RNNs, etc.
- Iterative optimizers such as CG, proximal methods, etc. Describe the tradeoffs involved, especially those which change significantly between CPU/GPU and FPGAs. Many of the techniques described in EE364B Lectures or CME302 would be appropriate for exploration. The primary motivation for these is for machine learning, so a motivation for the particular technique is necessary, along with a demonstration of it being applied (for the final report).
- Low-precision computing / cheaper approximations. This might involve replacing training-time SELUs with RELUs for inference, or sigmoids with piecewise linear functions. Existing nets could be modified to use FixedPoint or FlexPoint.
- Something else (propose a different idea).
Your project proposal should be submitted on Gradescope, due 11/6. It should include which of the options you are pursuing, and the goals of your project. As much of this work is squarely in the research domain, your grade will not be tied to how well you achieved these goals, but rather the degree to which your team explored and documented the area.
If attempting option 2 or 4, please also include a segment on the motivation for the work, as well as a summary of prior work.