In a quarter-long project (maximum 3 students per team), you will pursue a project dealing with Deep Learning accounting for 36% of the grade.
Please read and follow these instructions carefully.
The scope of the project is practically anything in deep learning. This could involve training a model for a new task, building a new dataset, improving deep models in some way, testing on standard benchmarks, etc. The teaching staff will work with you to ground the expectations and guide you as and when needed (please be proactive in asking for help). Finally, we also appreciate projects that solve personal pain points you face in everyday life or push the deep learning research further by a teeny-tiny amount.
We expect the project to be atleast 2.5 assignments worth of effort. If you want to leverage significant amount of pre-exisitng code, we would expect you to build something beyond (eg., a demo) on top of the base repository of pre-exisiting implementation.
At the conclusion of the project, your team will be responsible for writing a short research paper that summarizes the project. It should be 5-6 pages long (not including references) in the style of an IEEE CVPR conference paper (overleaf template, downloadable latex/word template).
If this is your first time writing a research project paper, here is a rough outline of sections that we recommend:
On the way to completing the research project, students will be required to submit a project proposal (1 page), and an intermediate project milestone (3-4 pages) report. Both of these will also be expected to utilize the CVPR template.
The project proposal should be a first draft of your introduction section of your final paper. It should try to answer the following questions:
The project milestone should be a draft of your final paper. It should include all sections except for the experiments section, which can be incomplete.
Each of these reports should be properly formatted, with the camera-ready version used in the overleaf CVPR template. Your project will culminate in final poster presentation. This is an in-person event meant to mimic a poster session in an academic conference. You will create a poster with relevant details from your project and discuss your posters with the attendees, which will be faculty and students from the department. The poster itself doesn't need to contain all the information from your final report, just the most important details to discuss the key points from your project in 5-10 minutes.
Q: Is novelty required for a good project?
No, see "How to choose a project". Novelty will lead to a strong project but outcome it's not required.
Q: Do I need to get state of the art performance?
No, most research contributions do not lead to state of the art performance.
Q: How else can I show effort?
Compare and contrast different methods, show multiple design iterations leading to improved accuracy, show creative design choices to tackle a new dataset. In general, thoroughness will show through as effort,
Q: How do I show proper analysis?
Do your best to answer “why” in your discussion! What kinds of mistakes is your model making? Where is it improving? Why does the loss/accuracy curve look the way it does? etc
Q: What do we do about GPU resources?
We expect that student groups will be provided $150 (for teams of 3) worth of Google Cloud credits to conduct experiments. See "Resources" for further details.