What is research?

  • a relevant problem – we have to care about the problem.
  • a tractable problem – we have to have some reasonable way of solving the problem (no time travel papers).
  • new and interesting information about a method you tried. Usually, this means a success, but negative results might be helpful, too.

Attention is All You Need is an example of a strong research paper. The team improved language modeling – a relevant problem because it is useful in translation, a tractable problem because we already had pretty good mechanisms for doing it (LSTMs). The paper shows successful results by using a new architecture called a ‘transformer.’

Finding a Research Project

When asked ‘where he gets his ideas from’, Neil Gaiman nominated ‘confluence’ as his best guess. He tried to consume a lot of content, then let it stew in his head until something new popped out. Academics are creative types, too, and we can use the same kind of method to find good research ideas. Here are my thoughts on signals to monitor in your search for a research topic.

friends

It’s a very good idea to develop relationships with smart, productive people in your field. You can collaborate and help each other.

newsletters

Some people make careers out of being information spreaders (and gatekeepers). If you trust the source, they can be a very helpful way for staying up to date.

Here are some I subscribe to:

conferences

Most interesting CS research is published as part of the ‘proceedings’ of an academic conference, rather than as a journal article, as in other fields. Although these conferences publish hundreds of papers each, you can glean the proceedings by developing your ability to speed-read abstracts and titles.

The conference websites are specific to a certain year, and usually allow you to browse their proceedings. You can pay special attention to the ‘best paper’ winners, not only because these are probably high quality, but also so that you can calibrate your sense for what makes a paper likely to be accepted.

Here are the conferences I’m most interested in, as an NLP / robotics person:

  • RSS
  • ICML
  • IJCAI
  • IROS
  • NeurIPS
  • CoRL
  • AAAI
  • ICRA

arxiv

Arxiv is an open publication site, mainly used by physics/math/CS people. Think of it as the wikipedia of academic journals – anyone can submit. However, there isn’t much in the way of a quality check, so be doubly skeptical of things you find on Arxiv (of course, this is a good rule with conference papers, too).

You can search Arxiv and find recent papers using keywords. As you develop as a researcher, you’ll naturally accumulate a few key terms you care about. But Arxiv is still very much a firehose, so you can use a tool like the arxiv sanity preserver to help you.

google scholar

Google Scholar is an excellent academic search engine, perhaps the best, despite the interface remaining unchanged for about a decade. You can pull the citation information for a publication, or look at the profile of a certain academic to see what they’ve written recently. (The academic might use it track their h-index…)

One of the handy features is to search for who has cited a certain article or publication. This lets you build a ‘snowball’ as you start with one very relevant paper, then develop a search tree from it.

the muse strikes during creation, not before

The best ideas aren’t the ones you start with, they’re the ones you come up with halfway through your paper. Try to resist the urge to keep casting about for the ‘perfect’ paper topic. You might have better luck by simply starting something silly and seeing what happens when you work on it.