Paper Presentation

Each student is responsible for presenting and leading the discussion for one or more research papers during the semester. Depending on enrollment, presentations may be done individually or in a pair. Presenters must email the instructor a shareable link to their slides at least 24 hours before class, such as Google Slides or PowerPoint on OneDrive, with commenting access enabled. When you are presenting a paper, you are not required to submit a review for that paper.

A primary goal of your presentation is to lead the class in a deeper analysis of the paper. Keep in mind that your audience has read the paper. Therefore, your presentation must go beyond merely summarizing the paper. You should provide additional insights and contexts, and foster discussions that deepen everyone’s understanding of the topic. Presentation that only reiterates the paper content will receive a minimum score.

Specifically, in addition to a brief recap of the paper’s motivation and main ideas, the presentation should:

  • Provide additional background knowledge not covered in the paper.
  • Highlight what points in the paper are important.
  • Explain technical details that may be overlooked or that are difficult to understand.
  • Use the optional readings to complement, compare, and contrast with the paper.
  • Analyze and critique the evaluation results.
  • Identify the strengths and weaknesses of the paper.
  • Describe your own reflections, including on potential future research directions.
  • Summarize key discussion points for the class.

Questions to Address for AI-for-Systems Papers

Because nearly every paper in this course couples a learning technique with a system, your presentation should also make the following explicit. These are the points where such papers most often hide their weaknesses:

  • Problem fit. Why is this systems problem a good target for learning at all? What would a strong hand-tuned heuristic achieve?
  • Where the model sits. Is inference on the critical path? What is the latency/memory budget, and how does the design meet it?
  • Data. Where does the training data come from? Is it available outside the authors' organization? How is distribution shift handled?
  • Failure semantics. What happens when the model is wrong? Is the system still correct and merely slower, or can it break?
  • Baselines. Is the comparison against the real state of the art, or against a weakened version of it?
  • Cost. What are the training, retraining, and serving costs, and are they accounted for in the reported wins?

The presentation should be interactive and include questions for the audience. You should also expect that the audience may interrupt your presentation and ask questions.

The presentation should last 20 to 25 minutes, followed by 15 minutes for discussion. As the discussion leader, you are responsible for keeping that discussion going. Prepare more questions than you think you will need.

The presentation will be graded based on the clarity, completeness, depth, and discussion. We will also ask the students in the class to rate each presentation and provide feedback. The audience rating is counted as part of your presentation grade.

If you present as a pair, we expect each member to contribute fairly to the presentation, and we expect to assign the same presentation grade to each member. To prevent free riding, there will be a peer-evaluation to assess each member’s contributions. Members who contribute less than their fair share will receive a lower grade.