Course Project

Teamwork

You are encouraged to form a team with 2-3 members, but you have the choice to work on your own for the project.
If you are not able to find a teammate, you can use Piazza (link) to create a post.
All teams need to finalize their team members by sending an email to Minxing Zhang (minxing.zhang@cispa.de) with all members' names listed in the email (due on Nov. 30, 2023). The email should use the title "Team - RML WS23".
In principle, each member receives the same score within the team, so everyone should be responsible for their team and put enough effort into their team's course project. It is your responsibility to distribute tasks and communicate effectively with your team members.

General Guidelines

The course project (60 pts) aims for you to explore a research problem in the field of robust machine learning.

You can choose any topic that is related to robust machine learning. The topics introduced during the lectures and the supplementary reading materials should provide you with enough hints about the research topic you can choose. You have the freedom to choose the dataset, the training algorithm, the model architecture, and even the robustness-related tool for your project.
The course project consists of an initial proposal (10 pts), an oral presentation (20 pts), and a final report (30 pts).

Below are example questions that may lead to a successful research project at different levels in pipeline of building robust machine learning systems:

Task-level: Are there ML robustness issues regarding the tasks/applications you are interested in? If yes, what are the specific differences/challenges compared with the robustness problem you learned during the class? Can you further formulate the definition of robustness by describing the attacker's goal, the attacker's constraint, and the attacker's knowledge?
Method-level: Which methods you have learnt in the class can potentially help address the robustness problem you defined? Can you adapt existing robustness training methods to solve the specified problem? Do you observe a trade-off phenomenon between your method's standard and robustness performance? Can you further improve the standard/robustness performance of those methods?
Evaluation-level: How do you systematically/rigorously evaluate the robustness performance of a given model for your task? Can you adapt existing evaluation methods for your task? How do you convince the practitioners that your method is useful for relevant real-world scenarios? Can you even provide an evaluation guideline for practitioners?
Explanation-level: Can you explain why your designed method works? Can you describe when your method will fail (e.g., different tasks, data distributions, machine learning methods or model architecture)? Does your method have limitations? If so, can you discuss potential ways to mitigate those limitations?

Proposal

Generally speaking, a research proposal should include the following three components:

A clear introduction to the robustness problem you plan to investigate and an explanation of why it is worth studying
A survey of related works on the problem you proposed
A high-level description of your method and research plan for tackling the proposed problem.

The proposal should be at most 2 pages long (excluding the references) using the NeurIPS 2023 LaTeX template. The proposal will be due on Dec. 21, 2023. You are encouraged to make use of the Q&A sessions on Dec. 14, 2023 to address your proposal-related questions with the TAs.

Presentation

The oral presentation will consist of a 10-minute presentation and a 2-minute Q&A. The time limits for presentations and Q&A may be adjusted later depending on the total number of teams. The oral presentation is scheduled at Lecture Hall 0.05, CISPA Main Building, between 13:00 and 17:00 on Feb. 08, 2024. If you have schedule conflicts for this scheduled time, please send an email to the instructor and the TAs as soon as possible. Depending on the demands, we plan to organize an additional time slot (TBD) in the following week for teams with unsolvable schedule conflicts.

Guidelines for Oral Presentation:

If you are working in a team, it is fine if one of your team members presents your project (on behalf of the whole team).
It is important to follow the time constraint strictly (10-minute presentation + 2-minute Q&A). We will have to halt your presentation if the presenter exceeds the limit.
The presentation should, at the minimum, convey the (a) motivation, (b) problem setting, and (c) your solution. I recommend around 3-4 minutes per part.
Be very selective in what you choose to present. For example, provide only the most compelling motivations or interesting results/insights. Do not simply dump everything that you’ve done.
Be extremely sparse on text. Do not dump a bunch of information in text on your slide. Instead, aim for a heading and, at most, one sentence per slide, accompanied by a graphic or plot.

Report

The report should be at most 8 pages long (excluding the references) using the NeurIPS 2023 LaTeX template. The report will be due by the end of Feb. 15, 2024. As for the expectations for the proposal, you should use the right template and follow the format requirements carefully.

Expected Structure for Final Report:

Proposal content. You should include an abstract for your work, an introduction, a comprehensive literature review of the related work, and a section describing the proposed work.
Summarize the work that you’ve done and what you’ve learned from it. Consider answering the following questions: For the work that you hoped to complete for the final report and did in fact complete: what was the result of the experiment or theory? Did it turn out fruitful, with interesting takeaways? If not, what factors led to this not producing an interesting result? For work that you hoped to complete for the final report and did not complete: what factors led to the work being incomplete? Were there underlying issues that you did not know about until now?
Conclusion. You should also include a conclusion for the final report. A conclusion that merely restates the entire paper is not useful. Great conclusions are inspiring and leave the reader wanting to investigate more, typically by (a) re-emphasizing any highly impactful results you have created or (b) showing the reader that there are still many interesting directions to explore. Conclusions are a great place to acknowledge the limitations of the current work as a future direction.
Reflections. This paragraph is not typical in a conference submission and is specific to the course project. The goal here is to think at a higher strategic level about your research process. Be sure to answer the following questions and reflect on how the project went: 1 • What did you find more difficult or easy than you expected? • Would you have done anything differently if you were to do the project again? • Where do you see this project going, hypothetically? If you had more time, would it be worth pushing certain directions further? Or has the initial research direction dried up and is no longer worth pursuing? It is easy to get caught up in the weeds of your work, and lose sight of the bigger picture. To be an efficient researcher, you can learn to reflect and optimize your own overall research process.

Use of Large Language Models (LLMs): You are welcome to use any tool that is suitable for preparing high-quality outcomes of the course project. However, you are responsible for the entire content of the submitted proposal and final report, including all text and figures. It is important to keep in mind that it is your responsibility to ensure the content of the submitted documents is correct and original.

Robustness in Machine Learning