Karl Stratos

A Note to a Prospective Student


First off, thanks for considering working with me. I apologize in advance if I'm not responsive. I lack the mental bandwidth to give the rightful attention to all that comes to me. Don't let my flaw discourage you.


Getting in touch

I've concluded that a litmus test for checking if we're a good team is for you to (1) first take my NLP course, (2) then do a project with me one-on-one, ideally in the summer when you can focus on research. The first step ensures you have the necessary knowledge, and the second ensures our standards and expectations are aligned (and also we get along with each other). If you're already at Rutgers, I strongly encourage doing this process instead of skipping a step: even though it takes more time, this will be more effective for both of us in a long run.

I understand that it's not always possible to do the process, especially if you're not at Rutgers. I don't mean to appear aloof! Please feel free to email me to ask anything about doing research. It'd be helpful if you explicitly tell me your relevant background in NLP and machine learning. If you're interested in applying to the PhD program, please specify NLP as the primary area of your interest and me as the faculty you're interested in working with.


Technical qualifications

I'm interested in unsupervised and lightly supervised (e.g., semi-supervised, transfer, multi-task) learning for NLP, with applications to knowledge-intensive tasks (e.g., entity linking). Doing research in this area involves identifying an interesting question, mathematically formalizing a problem, working out a solution, and then fitting suitable models on data to empirically solve the problem. It requires certain technical qualifications such as

  • Mathematical maturity.
  • A solid understanding of foundations of deep learning (e.g., gradient descent, backpropagation, variational objectives, noise contrastive estimation).
  • Superb coding skills. This doesn't mean knowing the syntax of PyTorch. It means general resourcefulness required to build a nontrivial system, like the ability to
    1. Abstract and decompose a large project into tiny pieces, and then solve each piece individually.
    2. Guarantee the correctness of an implementation by unit testing and constant vigilance.
    3. Use someone else's code by making minimal changes as necessary.
    4. Design well-thought-out experiments and methodically optimize key hyperparameters to maximize performance.
    5. Resolve any tedious technical issues (e.g., compilation errors) by mindfully reading system messages and searching the web.
You are not expected to have all of them from the beginning, but you are expected to work hard toward acquiring them. We will not be a good fit if we don't share an intrinsic motivation to grow in mathematical maturity, the knowledge of machine learning, and coding skills.


Thoughts on PhD


Now that we're done with the dry section above, I'll share some of my thoughts on PhD in a fireside chat form.


"PhD is a special time"

These were the very words that my own PhD advisor told me when I began my PhD. As I get older, I realize how true they are. Never before did, and never again will you have such a thick slice of time (5-6 years) you can dedicate to personal growth as a researcher under the aegis of your advisor who gives guidance when you need it and shields you from financial and administrative responsibilities. I went through metamorphosis in my time as a PhD student because of the sage mentorship provided to me by my mentors. I wish I can give it to you as well.


PhD can be a terrible time

In my class year, more than a half of the PhD students dropped out. Some went to industry and told me that they had been unhappy in the PhD program but were now much happier. Why? It's because there are so many traps that can trip you up in making your PhD a success. It's imperative that you make your PhD a success, not because it'll make you important but because you won't enjoy being a PhD student otherwise. What exactly is "success" here? To be blunt, it's all about publications. The two most important factors that define the success of your PhD are:

  1. Quality of your publications: you must be happy with your own work because it's substantial, fundamental, and original.
  2. Volume of your publications: you must publish regularly, otherwise you'll lack recognition for your hard work and lose motivation.
Among many relevant qualities, I'd like to highlight the following three as most important in being a successful PhD student: (1) self-motivation, (2) perseverance, and (3) focus.


1. Be self-motivated

Imagine two PhD students: John and Mary. John masterfully executes great projects given to him by his advisor, but rarely makes an effort to formulate problems and solutions on his own. Mary is less skilled than John, but she constantly brainstorms about what problems are interesting for what reasons and thinks deeply about technical solutions; she even brings these problems to her advisor and take ownership of her projects. In a short run, John's works may be better in quality and quantity because they're his advisor's ideas. But in a long run, it's Mary who will emerge as a successful researcher.

You must want to know the answer to your problem. You must want to publish works of great quality. What does it have to do with research? Everything! If you're not self-motivated, you will be a passive learner, and a passive learner is not a researcher. Advisors often encounter students who do exactly what they're told to do and nothing more. It's unlikely that they will continue research in a long run.


2. Persevere

I've seen brilliant students drop out of PhD because their first paper submissions are rejected. I've seen many projects that get abandoned half-done because initial experiments don't pan out. Doing research means most of your time is spent in limbo without tangible rewards. It also means you'll face opposition from people.

Stand your ground. Don't easily give up. Great papers are rejected all the time because the reviewing system is flawed. It will never be easy to take rejections, but they will come anyway so learn how to handle them. Expect the world to give you back less than a half of your effort, so be ready to put in a double amount of work if necessary. When your project doesn't go as well as you envisioned, (1) talk to people, (2) think hard about what's preventing progress and how to break through it, and (3) just try things. But also have the wisdom to recognize dead ends. You need indomitable perseverence and thoughtful flexibility at the same time.


3. Focus

Unfortunately, even when you're self-motivated and persistent, all your efforts can go down the drain if you lack focus. You must focus your energy like a laser beam. If you don't, you will be busy achieving nothing.

Many mistakenly believe that working on diverse projects is a mark of productivity. But it's not necessarily true. In fact, the only way to achieve true growth as a researcher is to single-mindedly drive a substantial project from beginning to completion. This is because your problem never reveals its true face until you get to the bottom of it. And you cannot get to the bottom of a problem without madly focusing on it for a very long time.


It's not about you

Please spend ten minutes to read Kevin Gimpel's advice on being a happier PhD student. I assure you it's ten minutes of your life well spent. I'll echo many of the same points, but a particularly important one is making your research not about yourself. In his words,

As long as you dwell inordinately on your brilliance and reputation, you will never be happy. Self-focus brings stress and kills the ability to enjoy the beauty of the discoveries you are making. Humility produces a receptivity to the wonder and richness of reality.

As soon as you make anything about yourself, you'll constantly fear disappointing others (about you), disappointing yourself (about you), and more generally failing anything (because it's about you). This is a miserable way to live. Forget yourself and "focus on the phenomena". You'll be struck with a realization that you can in fact enjoy your work because you can.


Don't self-promote

The world judges you based on how you compare with your peers in terms of visible achievements. Thus there is a toxic, and unscientific, culture of self-promotion: you have to somehow convince people that you're better than that person, while that person has to convince people that he's better than you.

Contrary to popular belief, I'd advise not to play this game. If you do this, you'll live in constant anxiety, either fearing losing the edge you have over others or feeling like a failure. The truth is, as mentioned in Kevin's advice: you're not the best, and you probably have a bug. Acknowledging this will free you from the unreasonable burden of having to outperform everyone and let you focus on actually improving your character, skills, and the quality of your work. And that's what this is all about. Let the quality of your papers promote you, not your tweets.


But you're brave

The vast majority of people in the world are not researchers. They do work that's given to them. In contrast, you're a researcher whose job is to identify work that's worth your time and doing it for the first time in human history. What does this mean?

It means you're brave. It means you have courage. Doing research means living with the fear of the unknown while feeling you're never good enough. As Nathan Schneider points out, in research

You have to first confront your ignorance on some question, then establish that the ignorance is universal. And you have to be willing to risk exploring new ideas that may go nowhere or may be ignored.

This is no easy work. Only a special kind of people can do it. Like you.


Serious means action

You show how serious you are by your action, not your talk. An action doesn't mean reading papers or debating ideas. It means really doing the painful part of work, such as banging your own head for days to identify a precise formalization of your problem instead of avoiding the pain by distracting yourself with other easier tasks and waiting for me or others to do this for you, or writing code that actually replicates baseline results.

Until you take an action, you're not really serious, and neither am I.


High-level understanding, without the details, is worthless

I'm often surprised by students who are convinced that they understand a method because they have a rough idea of how it works. This stage of understanding is so immature that they don't even have the capacity to understand their lack of understanding. The only way to understand a method is to know all the details. Not most of them, but all. You understand either everything or nothing.

And the only way to understand all the details is to create the method yourself. As Richard Feynman famously said, "What I cannot create I do not understand". You don't understand a mathematical theorem unless you can derive it from the premise. You don't understand an algorithm unless you can implement it correctly. Make it a habit to generate results instead of consuming them. Writing technical notes is enormously helpful (tips). Stop hypothesizing how you might improve a paper based on high-level understanding. You will never deepen your understanding that way. Start by replicating reported results.


Communicate frequently and honestly

Few things are as frustrating to advisors as radio silence. I've seen many students who abrubtly stop communicating, presumably because they're stuck in their project and/or they're not quite enjoying it. Please know that even if you do this you can always come back for help if you need it. But I'll consider an unannounced absence of communication longer than two weeks as a sign that you're not serious about the project, so I'll delegate it to other students.

Be honest about what you don't know. Communicate your thoughts often. This way, you will help me and also yourself.


View PhD as a formal business contract

This is an advice I wish I could give to myself during PhD. When the relation between an advisor and an advisee is formally established, it's like a formal business contract. The advisor's job is to provide critical guidance in research and career and also ensure that the student is protected from financial and administrative distractions. The student's job is to focus on producing research output, namely publications, in close collaboration with the advisor.

In PhD, you will have freedom to spend your time however you want. If you want to spend most of your time watching mindless TV shows and less than 10% on research, you can. But having the business contract model in mind is helpful in dispelling such use of time, since PhD is not supposed to be a waste of your time. It's supposed to be an intense, life-changing experience.

Here are some concrete applications of taking this view. First, prioritize producing research results above all else. Taking courses, TA-ing, and reading papers are all important and valuable activities to the extent that they help you in research. You must desperately resist the temptation to spend time doing these activities at the expense of your research output. Second, have a committed work schedule. For instance, when you're really onto something, have the passion to spend all your waking hours to drive the progress; but even when you're stuck, have the grits to work for 8 hours a day anyway. Having serious work discipline and being accountable will make you not only more productive but also more fulfilled (since you've "done your work") and therefore happier.


Be well

Get your life together. You can't do research if your life is falling apart because of serious health issues (physical or mental), addiction, or relational problems. I'm not saying that your life has to be perfect to do research: that will never happen (mine certainly isn't). But you have to optimize your well-being if you want to be effective in anything. Surround yourself with friends and family and love them. Have a routine—impose on yourself inviolable rules to find order in this chaos of life. Eat healthy and exercise regularly. Go to bed early, get at least eight hours of sleep, and wake up at the same time everyday. Each morning when you wake up, don't fill your mind with junk and worries by checking email or reading news. Instead, have an alone time and restore your perspective of life by meditating on good and truthful thoughts. Break the tyranny of this world by faith in God. When you need it, don't be afraid or ashamed of seeking professional help. Do what's necessary to be well while doing research.

© 2010–2024 Karl Stratos