Only one project proposal needs to be submitted per group. The idea behind this proposal is that I can give you feedback and suggestions to make you come up with a feasible plan for the project. You don’t have to do exactly what you say in your proposal, and can even completely change the project afterwards if you want. But it’s important to have at least one reasonable plan/baseline to start from. The length should be 2 to 4 pages, not including appendices.
It should contain title, authors, abstract, introduction, related work, model/method, and experiments. In particular, all these sections should be complete (though they can be short) except for model/method and experiments.
It should summarize the main idea of the project and its contributions. It should be understandable to anyone in the course.
It should clearly state the problem being addressed and why it is important.
It should clearly distinguish what your proposed contribution from previous literature (e.g., a 1-2 sentence summary of at least 3 closely related papers).
It should contain the description of your research idea. You can use equations or sketches to help explain the idea.
It should contain a list of to-dos (having some time estimation could be helpful) that could in principle make you finish the project. Here’s an example:
You should almost always start from simple ideas and gradually add the complexity since a research idea typically has to fail many times until it works. It is highly likely that the project doesn’t go as planned. Therefore, with simple ideas, you could have the flexibility to tweak and a higher chance of obtaining some conclusions in the end.
It should contain a list of to-do experiments (e.g., each experiment takes one subsection). One should clearly describe the tasks and datasets. One should also put placeholder figures and tables (with specifically designed axes, rows, columns, and so on). At last, one should write down the goal and expected outcome of specific experiments.
Typically, in machine learning and related areas, one should have at least one main experiment on a convincingly large real-world dataset to show that your proposed model/method empirically works better than prior work. Moreover, one usually adds other experiments on real-world datasets to analyze how important a specific model/method design is, i.e., the so-called ablation study. One can also design special synthetic experiments to showcase the benefits of the proposed model/method under certain (possibly extreme) conditions.