Write a Good Peer Review
Paragraph 1/5: Summary
- Summarize the paper in your own words, no quote.
- Be concise. Boil it down to the very essence.
- Be neutral. Do not yet talk about problems, concerns.
Paragraph 2/5: High Level Evaluation
This is the place for general comments that are subjective in nature:
- Did you find the paper well-written?
- Did the idea strike you as incremental or highly original? What does it build on / are the original parts?
- Were there sufficient experiments?
- Is the paper likely to be of interest to many people or of great interest to a small community?
Paragraph 3/5: High Level Technical
Some possible topics to consider are:
- Were any important parts missing?
- Does the approach make sense?
- Any evaluations that should have been included?
- What part excited you the most?
Don’t judge. For each criticism you have, give advice instead. Speak as if you are their advisors.
Paragraph 4/5: Low Level Technical
This is a place for very specific comments. It shows the authors you read the paper carefully. Most papers have at least one (small) mistake per page. For example:
- Any sections or equations that weren’t clear?
- Any figures that are hard to read? Captions of it given wrong? X tilde labeled incorrectly?
- Any mistakes with indices / equations?
Paragraph 5/5: Review Summary
This is a summary of your own review (not the paper). Do it in one or two sentences.
Make Good Slides
Overview
- Number of Slides: Allow roughly 1 minute per slide
- Sections: Problem statement, Approach, Results, Conclusion
Dos and Don’ts
- Never change notation throughout a talk. Keep notation consistent across figures
- Avoid appearance of only an equation. When you have an equation, also include illustrations of it to explain how it behaves. If that’s not possible, include some intuitive explanations.
- Label each figure very clearly. Don’t use unfinished figures.
- Always attach semantic meaning to symbols and animation effects and be consistent through the talk. e.g. If you want to use arrow → to represent “reduce to”, then use it with this meaning through the talk.
- Never assume people remember previous slides. Call back to meanings of symbols and other background information.
- Don’t assume your audience pays attention at all. If there is something important to say, say it repeatedly. Hammer the point home. Make it crystal clear.
Other Helpful Tips
- Each slide should have a single “take home” message. You should be able to verbalize that when asked.
- Try to stick to one running example. Explain every aspect of the paper in the same setting. (e.g. assume x is movies, y is movie ratings; x is images of faces, y is the age of the person; …)
- Use images, movies, etc (unlike this file) Don’t just do texts
Write a Good Research Paper
Layout
- Abstract: Summarize the paper in a few sentences (5-8)
- Introduction: Provide Motivation, explain why what you are doing is important
- Related Work: What have people done so far, what should the reader know about the literature
- Background: Summarize some tools that you use but your readers may not understand
- Method: Explain your method. Try to keep it as simple and clear as possible!
- Results: Evaluate your method. For each experiment be clear what point it is making, what are you demonstrating
- Conclusion/Discussion: What have you learned from your research. Where is this going? Any high level observations that are interesting?
Style
- Citations Cite generously. Always mention other work in positive light.
- Writing style NEVER try to make something look more complicated than it is. Always simplify!!
- Equations The more important an equation, the more text you should have around it
- Formality Avoid informal wording (“When we ran these jobs we got these results.”)
- Graphs / Tables When you discuss graphs / tables, say clearly what the reader should pay attention to. Don’t assume the reader draws the correct conclusion by themselves.
Explain Why, not What you did
What you did is meaningless without providing the reason. Avoid at all cost. E.g.
Now we integrate the function f(x,z) and obtain f(z). As a next step we add a constant term. The result is a loss function that we use for training.
You should explain why:
As our model is independent of x we can simplify our setup by marginalizing over x. The resulting function f(z) no longer depends on x, which allows us to ….
5 General Tips
- Do extensive background reading (you should know every paper ever published on the topic)
- Have a clear research question.
- Propose a solution and explain every step of the derivation. (Make it look like it is completely natural.) The more important an equation, the more text should be around it.
- Conduct clear experiments and make sure each one proves a well-defined point. (Be honest, you are researchers not salespeople!!!)
- Write well! Polish, polish, polish!