Final Project Proposal Guidelines

A google doc version of this document can be found here

Final projects often conform to one of the following options, though, of course, variation is always possible. Hopefully this at least helps you think through it.

1. Data Oriented: You are interested in articulating a new problem or a new research question that requires work in developing new data. For this type of project, you spend most of your time developing this dataset (maybe you want to address a novel application of an existing architecture to something that is salient to you and/or your community) and deploy existing models as baselines to show how good current models are at your task of interest.
2. Model Oriented: You are interested in solving an existing task and improving on existing models requiring work in hyperparameter tuning and model development. For this type of project, you spend most of your time improving your model on a fixed task in comparison to existing models (maybe you want to develop a model that generalizes better, uses less compute, draws on a different architecture) pushing the frontier on an established dataset.

Introduction/Motivation

• Broad Question
  • What is the big picture question you are trying to answer or problem you are trying to solve?
  • Why is this an interesting and worthwhile question or problem to work on?
• Narrow Question
  • What is the specific question or formulation of the problem you will pursue? (Note: you will need to pick something that is feasible to answer in 3-4 weeks)
• Outcomes
  • What are some possible concrete outcomes of your project? How do they bear on your question or problem?

Background/ Literature review

Find at least three papers related to your project. For each project write a paragraph or two summarizing: - What were the goals of that paper? How is it related to your project? - What methods did the paper use? - What were the conclusions?

Google scholar (or other comparable database search) is a better place to look than standard search: you are less likely to find blogposts with unverified content on Google scholar. Note, on Google Scholar you might see a lot of preprints from arxiv, even if they were also published elsewhere. It is good practice to try and find the most recent version of a paper. The general rule of thumb you should use: peer reviewed published papers are more credible than preprints.

Planned methods

• What methods do you plan to use? And how does it relate to prior work you described above?
• What kind of task is this? Supervervised, unsupervised using an encoder, decoder, etc.
• What will your data look like? Give an example of one (ideal) sample
• Is there an existing dataset you will use or will you have to create one?
• What type of model will you use?
• How will you evaluate your models?

Proposed timeline/ division of labor

Breakdown your project into separate tasks. For each task, list the expected amount of time, who plans to work on it and when they expect to complete it by. For this part, it might be most straightforward to fill out a table following the format below.

One of your tasks should be ‘’Prepare for pilot result presentation’’ and your timeline should highlight what work you hope to accomplish before the pilot results.