Posted: Monday, October 19. Due: Thursday, October 22 at 23:59.

This problem set is your project proposal. It turns your pre-proposal (Problem 3 of Problem Set 2) and the feedback from the October 13 session into a plan you can defend. Projects are done in teams of 2, and the proposal is one submission per team. Before writing, read the project page again and the pre-proposals of the other teams on #ddm_project.

The proposal should show that you have already started: the data is loaded, you have looked at it, and at least one baseline has run. Feel free to use a coding assistant to help you accomplish as much as possible in the coming days.

Here’s the Moodle Submission Link.

Format

Write it in LaTeX (Overleaf is fine), using a conference template (NeurIPS, ICML, or similar) so that it can grow into your final paper. It should be at most 3 pages excluding references. Put the title, the team members and their emails at the top. Every paper, dataset, library and model you build on must be cited in a references section (BibTeX, any consistent citation style), and cited in the text where you use it.

Writing a Proposal

The purpose of any proposal is to convince the reviewer to fund the project, whether the context is a research grant, a company initiative, or a course project. To do that, several elements must be clear and compelling. The structure below applies broadly to proposal writing and should guide how you organize yours.

1. Why Does the Project Matter?

Open by establishing the significance of your project. What are the applications and implications of the work? Why should a reader care? State the question you want to answer in one or two sentences, and a goal that is specific and measurable. Then be concrete about the impact: who would use the answer, what would change if it worked, and what we would learn if it fails.

2. What Is the Existing Work?

Show that you are familiar with the field and that you have read enough to know what you are talking about. Review at least 5 papers that relate directly to your problem, data or method: who did what, what methods they used, and what they found. Say how your project relates to them: what you take from them and what you do differently. To find 5 relevant papers you will skim many more; Google Scholar and reading abstracts first is the efficient way.

3. What Is Your Proposed Contribution?

After reviewing existing work, make the transition clear: despite all that has been done, there remains a gap. State your hypothesis and summarize your approach. Make your contribution explicit, and break it down into specific aims:

  • Aim 1: …
  • Aim 2: …
  • Aim 3: (if applicable)

Each aim should be concrete and actionable.

4. Data

Identify your dataset with a clear reference: its name, who collected it, a link, its license or terms of use, and a citation. Then describe it: how it was collected, its size, the variables (with units), the sampling in space and time, missing values, noise, and the preprocessing you need. Include at least one figure and a short table of statistics showing that you have loaded and explored it. If you generate your own data or combine several sources, explain how. This section should convince the reader that the project is feasible with this data, even if you change datasets later.

5. Models and Methodology

Propose at least 2 to 3 models, starting from a simple baseline and moving to the models you expect to work better (SINDy, symbolic regression, PINNs, neural ODEs, DMD, autoencoders, or others). For each, say why it suits your data and which scientific inductive biases it builds in: conservation laws, symmetries, known interaction terms, spatio-temporal structure. Report the training and validation metrics of at least one baseline you have already run. Then say how you will evaluate: the metrics you will report and why they fit the problem, how you split the data, and what you compare against. The key message is feasibility: the reviewer should finish this section believing the project can be achieved.

6. Timeline

Provide a week-by-week plan from now to the final presentation, with milestones for data, models, experiments and writing, and who does what. Say whether you need a GPU or AUB’s HPC cluster, and how much; start with a version that runs on a laptop.

7. References

Include a list of references cited in the proposal. This does not count toward the page limit.

LLM Conversation

You may use an LLM (ChatGPT, Claude, etc.) to navigate the literature and to push on your ideas, as you did with the PS2 agent. If you do, attach the conversation as a second PDF, or a link if your platform supports sharing. Use it to search for a real, meaningful contribution rather than accepting the first suggestion: ask critical questions, push back on generic answers, and refine your ideas. Using an LLM to produce a polished but shallow proposal will not score as well as using it to genuinely explore a problem space and arrive at a well-considered project.

Submission

Post the proposal as a PDF on the Slack channel #ddm_project by Thursday, October 22 at 23:59, one post per team, with both names in the message. Then read and comment on at least two other teams’ proposals. The proposal is graded on the clarity of the problem, the quality of the related work, how well the data is referenced and understood, how well the models are motivated, the baseline, and the realism of the timeline.