My 2022 Bloomberg Data Science Fellowship proposal can be found here.
Here are a few tips and tricks I learned along the way about writing a good fellowship proposal:
As opposed to a personal statement for grad school applications, I think you should treat the fellowship proposal as a “thesis proposal lite”. If you’re at the stage where you are applying for fellowships, you’ve probably got some research experience under your belt, an idea of a broader story you’re interested in pursuing, and hopefully at least one or two published papers that provide motivations for your pursuits. Pretend you’re talking to your thesis committee about what you want to work on, and convey those points here! Generally, I found this exercise to be extremely useful for my PhD process, since it forced me to step back and think about the broader implications of my research.
This is a style choice, but I try to start off papers and proposals boldly: with snappy, eye-catching statements about the world that grip the reader. In this proposal, I start off by saying that the language model pipeline is centralized. Make sure your statements about the world are not too wordy or long, and make sure the statements are not subjective enough to turn certain readers away. It is best if you can add citations to your first sentences. For the rest of the introduction, follow the general writing guidelines, e.g., use active voice. I sometimes like to end the introduction with an outline of the rest of the document, but that is not always necessary.
Your goal is not only to convince the sponsor that you have exciting ideas. You also want to show them that you have done previous work in your proposed research area that strongly suggests your ideas will be successful. And you want to show off that you can complete projects. So, all your ideas should build off previous work you have pursued. In my case, I chose to focus on modular LM architectures and decentralized LMs, which primarily built on our work on DEMix layers. In each section, my ideas for future work drew on specific findings from DEMix. Whenever possible, I tried to cite other relevant work that I published.
This means that no one will see your proposal outside of the funding committee, unless you post it publicly :). So, if you have that perfect sentence or two in one of your papers about something you’d like to argue, copy it over here! Obviously, you don’t want your proposal to re-hash one of your papers, but if there are some verbatim snippets, I think that’s totally fine.
There are many ways to go about organizing your proposal, and it’s probably going to be specific to the fellowship guidelines. For this proposal, I chose to focus on two broad ideas — one lower-level idea about modular architectures, and another higher-level idea around decentralized models but builds on the previous section. Bold keywords when possible to draw people’s attention.
However, make sure the proposal is cohesive with a single overarching idea. It’s tempting to talk about all the random ideas you have, but it’s better to convey one solid, well thought-out idea with lots of different sub-directions, than a maze of fifty ideas that may not be related to one another.
Remember that your research proposals should be interesting to the sponsor! Do a bit of research on the sponsor’s research or application areas, and try to tie in your directions to what they might be working on. For example, I chose to add a few sentences around how my work is applicable to Bloomberg’s interest in building models for breaking news, where domains are rapidly updating and modularity is useful. The Bloomberg fellowship also has a “social good” dimension, and so social themes weave throughout my document.
Feel free to add lots of future work in your proposal, including half-baked directions. For example, a lot of the decentralized LM section contains future work that I wasn’t sure would totally work, but had a vague idea about. These lofty ideas are the things that will get the reader excited to fund you. But again, make sure all your ideas are cohesive and supported by previous work.
I think writing a fellowship proposal was an amazing exercise to learn how to craft a longer term research agenda. Regardless of the outcome of your application, hopefully you’ll learn some new stuff about writing, generate novel ideas for research, and get a head start on your thesis!
Feel free to reach out if you have more questions, and if you are currently applying for fellowships: good luck!
]]>My personal research statement can be found here.
Disclaimer: I do NLP research and only applied to research programs in NLP. However, in a previous life, I applied to Neuroscience PhD programs, and these pieces of advice would have really helped my application back then. So I think they are pretty generalizable.
The two most important aspects to your graduate application are your recommendation letters and personal statement. My advisor, Noah Smith, has great advice on recommendation letters here, so you should check that out.
Here are a few things I learned throughout the process of writing the statement. I tried to highlight points that you can refer to my personal statement to understand further.
Your research statement should be opinionated about NLP. BERT and GPT-3 are great and all, but what are their limitations? Where do you want to see the field go from here? Where are the new frontiers of NLP, and why do you care about them? In my statement, I chose to highlight the brittleness and cost of NLP models, and how our evaluation methods don’t really transfer to the real-world. These topics wove together most of my previous projects, even those in neuroscience.
Tell a story. Lots of personal statements re-hash CVs; don’t do that. Instead, try to weave a narrative around your work, and if you have had the opportunity to do multiple research projects, see if you can draw connections and relationships between them. Selling your vision is as important as executing it. So this is good practice! Again, I chose to focus on one or two overall themes.
Just to emphasize this point further, make sure your ideas in your statement flow. Don’t start and end descriptions of new ideas without at least partially connecting them to previous ones.
Claim some ownership of area(s) of NLP that excite you, even if you haven’t delved into them a ton yet. Claiming ownership of a subarea doesn’t lock you into that area, but it does make you more focused and conveys scientific maturity. If you can describe projects you've worked on in your claimed areas, even better.
Even though I think you should be bold about your ideas, make sure to be humble; most of the projects we’ve done or witnessed in NLP so far are just initial steps towards our scientific dreams. Using the right language to convey appreciation for the complexities/difficulties of NLP problems will go a long way.
While you should tell a story, make it structured. Put yourself in the faculty members’ shoes; they are reading hundreds of applications each cycle. Get to your points quickly and in simple language. Highlight keywords if you’d like, and structure your essay in such a way that someone can understand your main points easily. I structured my essay with one paragraph per main research idea.
With respect to structure, there’s a ton of ways you could go. Here’s one structure that I like: try to start your statement off with your main themes. Pretend the application reviewer only has time to read the first paragraph. What do you want them to take away from the whole thesis? Then break the themes down into finer ideas, referencing previous or ongoing work when you can. To finish, zoom back out to your overall PhD goals.
Don’t be shy to use plain English, and always err on the side of simpler is better. While you want to get across that you know your stuff, you also want people to understand your ideas as easily and quickly as possible. For example, there was a lot of new technical ideas in the VAMPIRE paper, but I highlighted the most salient idea from the work in a couple sentences. The main idea I wrote about was that VAMPIRE is a cheap pretraining method using bag-of-words instead of sequences. Simple and easy to understand.
Feel free to discuss ongoing work. While I was working on my personal statement, we were submitting our eventual paper Don’t Stop Pretraining. I just talked about the current ideas I had around this project, even though they weren’t ironed out yet. This is a great thing to do; it will give you more to talk about during your visit days/interviews!
Make sure you mention who you want to work with at the university and why. Just 2-3 sentences towards the end are necessary, since faculty use this to direct your application to specific people.
Get people to read your statement I got probably 5-10 different people, some of whom were not super familiar with NLP, to read my statement and tell me if they understood what I was trying to convey. It was super helpful to get feedback on certain ideas that needed clarification or simplification.
Anyway, there’s probably a lot more stuff I could mention; I’ll update this post if I think of anything else. Feel free to reach out on Twitter if you have any specific questions.
Last thing I’ll mention is that there are a number of cool programs to get feedback on your application. I’d point you to the UW PAMS program where current UW CSE students are providing application guidance to prospective applicants.
Hope this helps, and good luck!
]]>