Confessions of a researchaholic

An excellent post by Sylvain Lefebvre on how to enforce rigor in research.

I assume you already know why you should read papers (if you are doing research). If not, ask your adviser(s) or senior members of your group. If they tell you that you do not need to read papers, or if you could manage to conduct research without reading papers, please drop me a note. I would love to learn how you manage that.

The primary goal of reading paper is to know “what” other people are doing, not “how” they do it (this is only the secondary goal). Mistaking the secondary as the primary is probably the most common misconception I have seen. It is true that eventually you will have to figure how the algorithms work, etc, but that is after you know what you want to do, the primary goal stated above.

For me, reading papers is like reading gossip or fashion magazines (I got plenty exposure to these thanks to my wife); I want to know what my friends and colleagues are up to, what the trends and future directions are, and what kind of topics would interest me as a potential research project.

As a corollary, try to maximize the marginal return of the time you spent on reading papers. Maybe there are people out there who are smart enough to be able to read all the papers, but for mortals like me, the suggestion is as follows. Instead of attempting to read the papers entirely (a common mistake by rookies), try to spend as little time as possible on each paper to pick up the gist or key ideas. For most papers, I usually spend only a few minutes (and sometimes seconds) going through only the abstract, introduction, images, and videos. The latter two are a blessing for graphics researchers; more often than not it is possible to know a paper by simply watching the video or flipping through the images. In particular, if the video (or the talk slides) is sufficiently informative, I could sometimes bypass the paper entirely. (So the best way to read a paper is actually not reading at all.) Plus, even for papers I have read entirely, I usually only remember the main points eventually.

Finally, if you visualize the distribution of papers you read in the ambient space of all papers, it should form a “T” shape, i.e. you should read board enough to cover all major topics in your field (the horizontal part of T), and you should be an expert on at least one subfield (the vertical part of T) that presumably will be where you publish.

After already collaborating (and publishing) with a double-digit number of students, I finally realized that I have been given answers and advices to the same set questions over and over again. So I think it is time to write them down.

For the benefit of my readers and myself, I will try to keep these posts as terse as possible without losing completeness (similar to the guideline for SIGGRAPH paper length). I will mark all my research advice posts under the advice tag.