from Guide to the PhD on Oct 16, 2022

How to write your personal statement, for PhD admissions

Most personal statements make for great origin stories — the prodigal 5 year-olds born to love deep reinforcement learning. Few, however, tell the stories PhD admissions committees are actually looking for: telltale signs of research savvy.

Like most PhD students, I've reviewed hundreds of applications for my university's PhD admissions committee. I've seen a large number of personal statements, both formally and informally. Fortunately, I've seen a number of great personal statements, better than my own. However, admissions is a crap shoot — there's a ton of luck involved, so your personal statement can't just be "good"; your statement has to be as pixel perfect as it can be — you're the opportunity of a lifetime and a PhD student no adviser will want to miss.

This post is written broadly for any applicant — you could have a modicum of research experience¹, tangentially related experience, or no experience at all. You could be an undergraduate, an industry veteran, or a master's student. Even for applicants with extensive research experience, I highly recommend reading this post. No matter the research experience, candidates have made all sorts of mistakes in their personal statement that could cost them invaluable reader attention. It's important to note that your recommendation letters are paramount in the application process, but the personal statement is the most controllable portion. Note these tips are written with the AI field in mind, where publishing is accessible and there are many conferences; the rules may differ in, say, mathematics PhD admissions. I will try to provide tips as general as possible.

In the section below, we'll start with the mindset of an application reader. This is the most important part, so you can understand the context for later suggestions and understand subtle tweaks I don't get to here.

Qualities we look for, in you.

Application readers are looking for candidates that would make great PhD students. This is an obvious end goal, but what makes a great PhD student is more nebulous. Here is a rough breakdown of the qualities we're trying to assess; I'll address several points you may have heard about admissions:

1. Show research readiness: Are you ready to conduct "high caliber" research? All of the below qualities are proxies for this. One obvious indication of readiness is to already be conducting research, published at known venues for your field. For AI admissions, you may have heard that top conference papers are necessary — NeurIPs, ICML, ICLR, CVPR, E/ICCV, ICRA, CORL, EMNLP etc. Unfortunately, given the number of applicants for AI admissions, having publications at these venues is almost an unvoiced requirement. However, you can still demonstrate research readiness without publications at these venues, or any venue for that matter. Below are several proxies for research readiness; I've certainly seen applicants get accepted by having one or more of these categories, in lieu of publications.
2. Deeply understand a problem space: Do you understand what an "interesting" problem is? This is often the basis for accepting applicants with industry experience, when the personal statement exemplifies this deep understanding. Notice I did say "a" problem space. It could be a problem space that academics are familiar with or, even more impressive, a problem space that academics are not actively working on but should be. For example, I knew one applicant who spent several years working on compiler optimizations at Tesla Autopilot; he subsequently had no problems applying to PhD programs, as he understood problems associated with real-world model deployments. Most importantly, this candidate brought a unique perspective on edge AI models that is difficult to obtain in academia.
3. Recognize and distill takeaways: Can you recognize an interesting insight? This is important for two reasons: (1) to pitch your own work and (2) to sift through large bodies of research, to find the nuggets you need for your work. The first facet of this quality is a bit simpler to assess: In your statement, how clearly do you communicate the insight for your own projects? This is the baseline: You should be able to communicate, at minimum, the technical insight that makes your work valuable. Given you've ideally spent many hours thinking about and working on those projects, these are the projects you should be best-equipped to distill into a few sentences. The second is difficult to convey and to assess in your statement, but the conclusion can be utilized for this. We'll discuss specific examples of this later on.

There are 3 meta-points to make about your personal statement. These aren't qualities we're looking for per se, but you'll need these properties to make your case:

• Write clearly: Your writing should be clear and concise. This sounds like a no-brainer written this way, but there are many personal statements that look like first drafts. This means run-on sentences, grammatical mistakes, and unfamiliar technical jargon. Perfect grammar is not necessary per se, and none of these are reasons to dismiss an applicant. However, the statement needs to be readable. Lack of readability certainly makes the reader's job a lot harder — this in turn makes the reader harder to convince.
• Communicate ideas concisely: The reader is assessing both your research and your ability to communicate it. At the end of the day, assessing both is simply whether or not they're convinced of the importance and insightfulness of your work. This generally means communicating the vision for your work — the impact it has and the higher-order problem it solves. It doesn't mean embellishing.

At the end of the day, the PhD admissions committee is looking for researchers to invest in. They aren't picking applicants for their research agenda, for example, so really focus on exemplifying these qualities. Notice none of these qualities require publications per se; publications are, like most other things, just a proxy for these qualities that predict your future trajectory as a PhD student.

How we (try to) identify these qualities.

It's hard to assess and catch all nuances that justify your research readiness, so your goal is not just to state proof. Your goal is to make it extra extra clear that you are well-equipped to conduct high caliber research. In this section, I'll describe how I read the personal statement, so you can get an idea for where your reader's attention could falter or be reinvigorated:

1. Read the introduction to see if I should continue reading. I skim the introduction to find evidence of readiness: publications, experience, etc.

   1. If I don't find clear statements of these, I'll skim the rest of the personal statement to find these indicators. Assuming I don't find any, I'll mark down that the personal statement does not add new perspectives or context to the application.
   2. If I do find clear statements of this evidence, I make a quick note of these. I'll then check them off, with extra notes as I proceed through the statement.
   3. Experience is relevant background for the field. For example, if applying for AI-CV (Computer Vision), the applicant could have spent significant amounts of time working with real-estate captures from professional and handheld cameras. In this case, research readiness comes from a working understanding of photos used in real-world applications. We value perspectives we don't have.

2. For each project in the statement, I'll look for key parts of the "research story": problem, insight, and impact. Often times, the project description is a jumble of facts that partially or don't address these pieces. If I spot a familiar venue (e.g., one of the conferences listed above), then I'll let a less-clear statement of the problem or insight pass. In one scenario, I saw a paper that I had previously cited and read! This is of course very rare, but this left a very strong impression that I still remember today. Even (and maybe especially) in the absence of a known venue, always include the problem statement, technical insight, and quantifiable impact for your projects:

   1. Look for a problem statement. Is it interesting? Important? Is it even mentioned and clearly explained? Often, the paragraph jumps straight to the solution. For example, "we applied a 3D convolutional neural network to detect tumors". This is completely uninteresting. We'll talk about ways to fix this later on. At this point, I'm trying to assess the significance of the project.
   2. Once I understand the problem, I'll look for the main insight. Does it teach me something new? How is it used to solve the problem? At this point, I'm looking for cleverness that I can appreciate and remember. If so, I'll note that down.
   3. After the problem and insight, I'll look for results and impact. The result is ideally quantified in some form — an accuracy number or percentage latency reduction for example. Impact is a bonus here — it could be the new applications it enables, or in very rare cases, a real-world application it benefits.
   4. It's worth re-emphasizing here that the project doesn't need to be published to be interesting. A recognizable venue is (sometimes) an indicator that the project is already interesting, certified by another researcher in the field. However, there are certainly statements that do a good job of discussing non-published work, by addressing the points in the process above.

3. For the conclusion, I'll look for a description of interests. Are they interested in sensible problems? I again look for problem and insight. For obvious reasons, I'm no longer look for results but a statement of the impact it could have, is important. At this point, provided I'm convinced the applicant is research-ready, I'm looking for reasons to be excited about their future. I'll make a note of their future research direction. Added bonus if I happen to think this candidate would be a great fit for my adviser's group. In which case, I'll also ping my adviser to let him know to check out this application.

The process above is pretty barebones. There are specific items I'm looking for, and the rest of this post will discuss mistakes and recommendations that optimize for these items.

Always include the problem statement, technical insight, and quantifiable impact for your projects.

Top 3 mistakes: Origin stories, non-problems, what instead of why.

Here are the top 3 mistakes in personal statements. Without sugar coating, these mistakes waste space by addressing none of the qualities readers are looking for.

To avoid all of these mistakes and more, do the following: For every piece of your personal statement, ask yourself: does it show research readiness? Here are the top mistakes that fail this criterion:

• Tells an origin story. This is more common than you might think. Many statements paint the picture of a child prodigy, born to love mathematics and fascinated by science at the tender age of 3. There are loads of irrelevant stories — for example, how the applicant saw a baseball match as a toddler and wondered why the ball didn't fly forever after being hit. Needless to say, this is completely irrelevant; it tells me nothing about your research readiness today. There are varying degrees of the "origin story" and some are more relevant than others. Even if you're discussing completely relevant experience, avoid storytelling — at least, beyond the 3 elements we discussed above: problem, insight, and impact. Anything not in one of those 3 categories is fluff.
• Missing the problem. Given the discussion so far, this mistake seems obvious. However, it happens often. Many times, project descriptions will lead with and never proceed beyond the solution. For example, "We generated dog barks using a generative adversarial network". The paragraph would then proceed to describe details for data collection, method, or tell a story. Instead, the problem needs to take center stage. Your problem description should explain (a) why this problem is challenging to solve and (b) why this problem is important to solve. For example, "Generative adversarial networks (GANs) have been shown to work on large datasets, mainly of natural images, but such large datasets don't exist for dog barks". This shows difficulty. Next, explain the technical importance. For example, "This is a problem for GANs, which need large quantities of data to converge; even in this sample-abundant scenario, GANs often fail to converge, making small datasets especially difficult to train on". This hints at the importance of the problem: If we tackle this problem, we may solve a broader technical problem with this type of model. We'll work through more examples in the next section.
• Describes what not why. This is in effect to say that many project descriptions miss the problem and insight. The description simply reads like a methods section of a paper: "We collected data from 16 dog breeds in the Los Angeles area. The data was then collated into a dataset using a 70-30 training-validation split. Using PyTorch, we fine-tuned a pre-trained model…". You can imagine the rest of this project description. It focuses too much on the what part of the project. Instead, focus on the why. The problem addresses several higher-level questions: Why are we working on this? Why is this not a trivial problem? The insight highlights the cleverness in your approach: Why has no one else realized this obvious solution? Why does this insight solve your problem?

For every piece of your personal statement, ask yourself: does it show research readiness?

The most important takeaway is to avoid any sentences that do not suggest research readiness. Get to the point and do so concisely. Next, we'll discuss concrete tips and walk through examples.

Tips for writing your personal statement.

There are effectively 3 sections to your personal statement: the introduction, summarizing your research readiness; project descriptions; and the outro, expressing your research interests for future work. Below, I'll describe some general tips for writing each of these sections. For a collection of online personal statements, see Alex Lang's collection for the NSF GRFP, where you can judge how well each statement leaves an impression on you.

Introduction. The introduction is fairly simple. It simply summarizes all the facets of your research readiness. The most important part is to summarize the rest of the statement. There are 4 categories of topics to include at this stage, at least in the successful statements I've seen:

• Motivation: Don't spend too much time on this, but a sentence is acceptable. For example "My own visual impairments led me to study computer vision for accessibility". This doubles as proof of deep problem understanding, as you know which problems are important to solve for the visually impaired.
• Research experience: If you've published papers, mention the number of publications, where they were published at, the type of publication and your authorship. If you have space, briefly mention the topics for these publications e.g., "few shot object detection".
• Related experience: Mention any experience you've had that is relevant. Continuing our music example from earlier, say you've been writing music for several years; it'd be worth mentioning that your music creation hobbies taught you which biases to imbue into the models you're training. This should be brief, not more than a sentence or two. Other examples of related experience including giving technical talks, attending conferences, or deploying a product related to your field of study. Note that these are useful to include only as evidence that you're actively engaged in the research community; they aren't themselves accomplishments.
• Problem: If your projects all address a common theme, the introduction would be a great place to motivate the general thesis for your work. It's highly unlikely you'll have a central thesis, but if you do, this is one way to organize your introduction.

Your problem description should explain (a) why this problem is challenging to solve and (b) why this problem is important to solve.

Project. Each subsequent paragraph should effectively cover a new project. Every project should in turn feature the 3 parts of a problem that we discussed earlier: problem statement, technical insight, and quantifiable impact. Note that even if your project is not ultimately published, you can still use it to your advantage by following these guidelines. Making these parts clear is more important than the status of your project, as you ultimately need to convince your reader that your work is important.

• Problem statement: Describe the problem your project addresses. First, make sure to cover why the problem is challenging. Why has no one else successfully tackled this problem? This may be due to lack of data or inapplicability of state-of-the-art methods to this domain. Second, make sure to cover why the problem is important. Why should the reader care about problem? This could be due to a flaw in the field as a whole or influence how all models of this type are built in the future. Note this should be a technical importance. Any broader impact can be saved for the last section.
• Technical Insight: Describe the insight that led to your method. This is distinct from the method itself; it's the key idea that other researchers have missed, that you recognized and capitalized on. What is the one tidbit of knowledge that is so obvious in retrospect, that you needed to develop your method? You should be able to follow this statement with "Using this insight, we developed…". Of the 3 critical components listed here, this insight category is the most "skippable". A strong problem statement and quantifiable impact can certainly make up for missing insight.
• Quantifiable Impact: Ideally quantify results for your work, and explain the significance. Quantifying results is straightforward in theory; in practice, it can be difficult. Find anything you can to quantify: Show accuracy wins, latency reductions, developer time saved. Anything. Significance here is different from problem importance. Instead, this is broader impact, beyond just technical impact. Broader impact could include enabling developer productivity or catapulting a field of research into a new, previously-overlooked category of problems.

Here are a few more notes about project paragraphs:

• For transitions between paragraphs, unless it falls in the "problem" category, avoid more than a sentence or two of transition. One reason to linger would be if one project uncovered the motivation for the second project.
• As I live under a rock, I rarely know the faculty listed in the project. You can list the faculty member to give the work some ethos, but there's no need to explicitly list every collaborator and PhD student on the project. Some particular faculty are of course very famous and worth mentioning but generally speaking, give less importance to names.

Outro. Your final paragraph should indicate your future research interests. Express interest in PhD-student activities, most definitely in research and possibly in teaching, mentorship, presenting technical content, and service to the research community.

Conclusion

There are certainly exceptions to the guidelines I've provided above. There are other fields that may be more interested in the origin stories. There may also be applicants that wrote a wishy washy personal statement and were still accepted; I would argue these are the exceptions and not the rule. The above tips gear your personal statement up to be the most efficient at communicating your work these past few years. If anything, it's worth putting in the extra time now, as an extra few hours now will matter disproportionately much.

You can find my own personal statement, which I used for both my fellowship applications² and for PhD applications in general, at this webpage with fellowship application tips.

← back to Guide to the PhD