Humble Advice on Applying to Fellowships

Every year I am asked for some advice about applying for graduate or postdoctoral fellowships – usually regarding fellowships I have been lucky enough to hold (NSF GRFP, Microsoft PhD Fellowship, Miller Fellowship). Usually, these requests come from graduate students who are quite close to me in social network distance – corollarily, they are usually already members of the small set of academic institutions I have been affiliated with. All of this seems to perpetuate a “rich get richer” phenomenon in academic fellowships (counterbalancing this, of course, is the fact that my advice is probably pretty useless!).

As a small step to counter this, I am publishing a number of example research statements that I used to apply to these fellowships, since I usually send them to students who ask to see them. Someday, I will write some platitudes about how to write fellowship applications and research statements. But I do not have that much time, so for now I am also putting below a lightly-edited email conversation I had with a student about the Miller fellowship application process.

Finally, in the spirit of the CV of failures, let me list some fellowships and awards I applied for and did not get:

• 2017: Simons-Berkeley Fellowship
• 2017: NSF Postdoctoral Fellowship
• 2017: Harvard Society of Fellows
• 2018: ACM Doctoral Dissertation Award

Statements

NSF GRFP: Personal Previous Research Research Proposal (Note that the GRFP now requires a somewhat different set of statements.)

Microsoft PhD Fellowship: Research Statement

Miller Postdoctoral Fellowship: Research Statement

Miller-Specific Advice

A thoughtful student sent me the following questions about the Miller fellowship research statement:

1. How much of the emphasis should be on prior work versus future work?
2. To what extent do they care about a research plan having broader societal impact versus having technical depth/focus? (not that these are necessarily mutually exclusive)
3. Should I spend part of the statement mentioning specific people at Berkeley, in addition to my faculty host, that I would want to collaborate with and on what problems?
4. When they say it should be tailored to a broad scientific audience, is it safe to assume some basic familiarity with ML/statistics, say, the notion of gradient descent, or of learning the underlying parameters of a distribution from samples?

Here is my reply (lightly edited):

It is hard to give very concrete advice because (a) I have a sample size of 1, more or less and (2) the executive committee is never very clear about how they actually choose fellows, even when talking to existing fellows. But with those caveats:

The “broad scientific audience” part is very important — probably your statement will be read by a biologist, physicist, etc.. I think the statement should make it clear that you have a compelling research agenda of your own, have the technical capabilities to solve the hard problems in that agenda (as evidenced by your past successes), and that some success is achievable in the 3 year span of a postdoc.

1. I think this is up to you…I spent about 1/2 on broad agenda, 1/4 on prior work, and 1/4 on future work.
2. I think they care that you are doing important and fundamental science, whatever that means. It is definitely not a requirement that your work have immediate societal impact; the institute is very theory friendly. It is called miller institute for basic research in the sciences and they take that quite seriously. You definitely do not have to spend a lot of space motivating your work with the latest neural net self-driving car etc kind of applications. That said, it still needs to be crystal clear to a non-CS person why what you’re doing is important and fundamental. You don’t necessarily have to motivate things with societal impact, but you do need to explain why you care about them.
3. Probably a few people. I think they are looking for evidence that you have several shots at successful collaboration.
4. Gradient descent is probably ok. Learning parameters of a distribution from samples sounds a little too jargon-y. I would say something like “fitting a model to data” (which I realize is not exactly the same thing).