Each week you are required to read one pair of papers (either the 2 Tuesday papers or the 2 Thursday papers). At least 15 minutes before the start of class on the day that the 2 papers you selected will be presented, you should post on Piazza, in a private message to the instructors, a review of the papers. Your paper reviews should consist of 3 sections:
1 section summarizing the first paper, using the format shown below
1 section summarizing the second paper, in your own words (do not copy text from the paper)
1 sections on any connections you see between the papers. This is open-ended, but some topics you can discuss:
Compare and contrast the approaches.
What are the limitations of each approach?
Propose a way to combine the two papers.
Propose a new idea to address some limitations of either of these papers.
etc
Your reviews should be submitted via Gradescope.
For the sections summarizing the papers, please answer these questions (each will be a separate question on Gradescope):
What task is being solved?
What are the inputs to the learned component (e.g. the inputs to the network)?
What is the output of the learned component (e.g. the outputs of the network)?
What training signal is used? For example, if RL, what is reward? If not, where does the ground-truth for training come from? What is the loss?
What insights did they use to enable their method to successfully learn the policy?
For each section, you will receive at most 2 points:
Paper 1 summary: 2/5 point for each question
Paper 2 summary: 2/5 point for each question
Comparison paragraph: 2 points
For each answer, the following fraction of points will be awarded:
0%: answer is missing
50%: answer contains little information or mostly misses the point
100%: answer demonstrates a clear understanding of the paper(s)
You still need to turn in a paper review on the weeks that you present a paper. If you wish, you can choose to submit a review of the 2 papers on the day that you will be presenting (submitted 15 minutes before the start of class, as usual).
Late reviews will receive a maximum of 1 point per section, so please submit your reviews on time!
For extra credit, you can review all 4 papers in a given week. We will grade these extra reviews and add them to your reviewing score for the week.
Q: "The paper summary format should be flexible. Since in many papers, the most important focus point is not input/output or the way that networks are trained. Maybe a better idea is to summarize the paper and write the major contribution of the paper as well as its pros and cons."
A: We previously had a more flexible format for paper reviews, but students would often just copy-paste the abstract (perhaps with small modifications) and submit that as their paper review, without actually reading the paper. Thus we added more specific questions to make sure you actually read and think about the paper. The 5th question for the current paper review format is very general: "What insights did they use to enable their method to successfully learn the policy?" as well as the last question: "Describe the connections you see between the two papers." The questions that you suggested are somewhat subsumed by those two questions.
Q: Can you release the discussion questions earlier, so that we can think about them while reading the papers?
A: We cannot; we allow students to submit discussion questions at midnight (EST) the day before class, and then the instructors need time to read, filter, and edit these questions into the final set of questions that are used to guide the class discussion.