CMSC 35200 Project and Paper¶
The project and paper are an important part of the course. You have to define and complete a project that covers some aspect of deep learning systems. Based on your project, you have to write a final paper evaluating its features and performance.
Students may work alone or in teams of two or three.
Schedule (subject to change)
October 17: Initial project proposal due–one page writeup of project goal, research plan, proposed outcome(s). Your instructors will provide feedback.
October 24: Final project proposal due, refined based on feedback.
November 14: Mid-quarter project report. 2-3 pages summarizing progress, any problems encountered, and any changes to project direction or scope.
December 1 and 3: Project presentations
December 8: Papers due. 8-10 pages in ACM Conference Proceedings format (sigconf).
Some Potential Project Ideas
We encourage students to propose their own ideas for projects. Here are a few ideas.
DNN training and inference involve many vector and matrix operations. The performance achieved on a particular cocomputer thus depends critically on whether vector and matrix sizes are well aligned with the computer’s memory architecture. A potential project would investigate how these factors effect performance on different computers, and how DNNs might be modified to increase performance.
Nicolas Masse and colleagues propose neuroscience-inspired approaches to avoid “catastrophic forgetting” in neural networks. A potential project might apply these techniques in other contexts and evaluate their performance.
Implement, evaluate, and document programs on Cerebras, SambaNova, and/or GraphCore