30K Subscriber Special: The Gradient Editors on Grad School

Some of our Gradient editors on their experiences in grad school!

Hi friends! The summer is coming to an end (depending on where you live) and we’re roughly in grad school application season. We recently reached an unbelievable milestone of 30,000 subscribers—thank you all so much!—and I thought some perspectives from people who’ve dealt with grad school might be helpful to a lot of you.

So, in the midst of GPT-Everything, ElonTwitter, and who knows what else is going on in the world right now, I took the opportunity to bother a few of our wonderful editors who are trudging through PhD programs right now about their experiences.

If you’ve got more questions on applying to grad school, picking schools/advisors, or anyone else, they’d be more than happy to answer your questions! You can leave them on this post or join us in our Substack Chat I’ll make to accompany this article.

Self Intros

Derek: I’m a PhD student at MIT CSAIL working on learning methods that meaningfully account for the structure present in data. I’m currently working on equivariant (symmetry-preserving) neural networks and graph neural networks, with a mix of theory and practice.

Hugh: I’m a PhD student at the Harvard EconCS group working on game theory and multi-agent reinforcement learning. I’m very excited about bridging the gap between game theory and reinforcement learning and creating “general” learning agents that can simultaneously learn in both environments.

Kiran: I’m a PhD candidate at the Diagnostic Image Analysis Group (DIAG) of Radboud University Medical Center. I work with deep learning algorithms to detect lung cancers at an early stage from chest CT scans.

Andrey: I received my PhD from the Stanford Vision and Learning Lab in 2023 working at the intersection of robotics and computer vision, where I was advised by Silvio Savarese. My focus is on the broad set of problems related to embodied object search, including manipulation, semantic reasoning, and memory of past experiences. I now work at Astrocade, a generative AI startup.

Q&A

What motivated you to start grad school? What was your experience with the process? 

Derek: In undergrad, I found that I really liked learning about topics in computer science and math, and I quite liked research as well after trying it out. Also, machine learning advances were getting more and more exciting towards the end of my undergrad (e.g. generative models like GPT-3 and DALL-E), so I really wanted to do work at the cutting edge of the field. I applied for PhD programs in the last year of undergrad, and I directly joined MIT for my PhD after graduating undergrad.

Hugh: I felt that AI was important and felt like grad school was a good place to learn more about it. As a former Go player, I first learned about AI when I saw Alpha Go beat Lee Sedol, the former world champion. After that, I studied economics / AI as an undergrad and my experience with game theory led to a natural segue to multi-agent reinforcement learning.

Kiran: I worked on denoising autoencoders analysing brain MR images during my master’s. After my master’s, I joined a startup to build deep learning algorithms for medical image analysis, and over time, I realized that clinical products not only required very accurate algorithms but also required peer-reviewed publications to build scientific trust among clinicians. After three years at the startup, life happened, and I decided to look for jobs in Europe. During the job hunt, I came across PhD vacancies at DIAG and decided to apply. I’d already come across DIAG in my previous roles because of grand-challenge.org – a platform for hosting and participating in challenges in medical image analysis – and it was an easy decision for me to join them.

Andrey: I first got into research as an undergrad after finding my intro to AI class really interesting. I saw a poster for CMU’s Robotics Institute Summer Scholars summer research program somewhere and applied on a whim, and wound up doing that instead of an industry internship. I enjoyed that, so I then proceeded to do research as an undergraduate in one of my college’s robotics labs for the last two years of my studies. I tried working in industry for a bit after undergrad but got bored, and so decided to apply for a Masters to do more research and decide whether to pursue a PhD or not. I once again found research exciting and cool, so decided to go for it.

---

What’s your research area and what interests you most about it? 

Derek: I work on incorporating the structure present in various types of data into neural networks. I do a mix of theory and application, and work on methods that are hopefully applicable to many things at once.

Hugh: Multi-agent reinforcement learning. RL has taken a backseat to language models in recent years, but I’m hoping that MARL will help it make a comeback.

Kiran: I work on AI for lung cancer screening. This involves building deep learning models for detecting early-stage lung cancers from chest CT scans. These scans produce huge 3D volumes of the insides of the chest, and early-stage lung cancers can be pretty small (sometimes less than a centimeter in size), hard to detect and grow rapidly over time. My task involves tracking these nodules over multiple CT scans and designing deep learning algorithms to differentiate non-cancers from cancers accurately.

Andrey: I work at the intersection of robotics and machine learning. Most of my research has focused on different aspects of embodied object search (enabling a robot to find a specific project), which is a useful problem to solve with plenty of interesting challenges. My research has addressed a variety of these challenges, ranging from perception, physical manipulation, semantic reasoning, planning, and online learning.

---

What’s difficult for you about grad school? What are some of the hardest lessons you’ve learned?

Derek: It has been hard to balance short term vs. long term planning. For instance, some research ideas are safer, and you can make technical contributions through these ideas quicker. On the other hand, some potentially more impactful research can be riskier and may not work out, but may have long term impact if it does. Also, things like reading textbooks and getting good fundamentals are long term investments, but this can be hard to balance with research.

Hugh: Opportunity cost on what to work on. There are so many interesting problems, but you can only work on one direction when all is said and done (or at least, one direction at any given moment). The balance of choosing a topic and having to see it through versus being tempted on all the other interesting problems is a tough choice.

Kiran: Research can be hard. It can be a lonely experience if you are stuck in a rabbit hole. But that’s also where you get to build independent thinking and learn to be on your own and pull yourself out of the rabbit hole - through collaboration (with colleagues, friends, and family, and also with yourself).

Andrey: I wrote a post called Lessons Learned the Hard Way in Grad School which I think is still fairly accurate. In short, research is hard — there’s lots of uncertainty, lots of dead ends, lots of confusion. There are plenty of lessons about how to best do research you’ll only really understand through struggle and failure. The hardest lesson is probably just how to not compare yourself to the brillaint people around you and how to not feel bad when progress is slow.

---

What does it mean to you to do good research?

Derek: One of the major goals in my research is to generate and explain ideas that make people think differently or understand certain things better. This can be done through making general frameworks that generalize existing works, developing theory to understand certain things in machine learning, or making improved models that are different from popular models. Another sufficient condition for good research, which often makes for more short-term applicable research, is a solid technical contribution towards things that people care about, e.g. improving efficiency of some algorithm that researchers care about.

Kiran: Good research involves a thorough investigation into a particular topic, formulating clear research questions, collecting and analyzing data, and critically evaluating and interpreting results. Good research also requires clear communication of the insights you have gained to a broader audience.

Andrey: Good research results in new and useful knowledge, which is communicated in such a way that other academics and practitioners can make use and build on top of.

---

What’s something you’ve learned about yourself in grad school?

Hugh: Healthy habits are more important than I thought, and I already thought they were pretty important.

Kiran: It’s very much possible to translate skillsets across academia and industry. In academia, you will also have to solve poorly defined problems while collaborating with multi-disciplinary teams.

Andrey: I’m not quite as skilled a multi-tasker as I initially thought; learning to focus on a single research project really helped me do better.

---

What’s a paper or direction in your research area that you’re excited about? What’s one you feel is underrated or deserves more attention?

Derek: The analysis of symmetries within neural network function classes, loss landscapes, and training dynamics is very exciting to me. I think further work in this area could greatly improve our understanding of neural networks, and lead to advances that improve neural networks in the future.

Kiran: The self-configuring nnU-Net. It won the medical image decathlon - a challenge that involved accurately segmenting regions of interest in ten different tasks. It’s a method based on the simple U-Net – without the bells and whistles which have often plagued the field – that won the challenge purely because of its ability to systematically configure its architecture and training procedures by understanding a dataset’s “fingerprint”. Along the same lines, I believe that data-centric AI is an underexplored research area.

Andrey: The recent work from Google AI on using LLMs for robotics (Socratic Models, Code as policies, internal monologue) is really exciting. The ‘common sense’ needed to decide what to do in a given situation is one of the most challenging aspects of robotics, and these works point in an exciting direction.

---

What motivates you as a person?

Derek: After completing some research, I always like to think about how this work may improve some technology some years into the future. I am quite motivated by improving technology that improves society.

Kiran: Creating clinical impact with technology.

Andrey: Doing things I think are exciting or cool.

---

What are some things you admire about your fellow PhD students / advisor?

Derek: I just love their mentality and the day-to-day motivations of their work. They are so driven by intellectual curiosity, and thus they each have such a breadth of diverse knowledge, ranging from esoteric mathematical subjects, to deep understanding of particular algorithms, to an amazing memory of different empirical tricks and hyperparameter choices used in a field within the last two years, to knowledge about algorithms used for very specific problems in computational biology.

Kiran: Every PhD candidate that I’ve known has evolved into a very mature researcher. And, my advisors have had a excellent track record of taking research solutions into clinical practice - something that made me want to work at DIAG in the first place.

Hugh: My advisor works so hard and is so kind. I aspire to be like him in these regards.

Andrey: I admire fellow PhD students who go out of their way to contribute to the community in ways that not usually encourages, such as mentoring students, hosting social events, communicating science, hosting reading groups, etc.

---

What advice would you have for…

Someone considering grad school?

Kiran: The experience that you gain from applying the scientific method on a long-term project, where you are also given time (typically 4+ years), is definitely worth having. Doing a PhD also opens the path for you to create your own research lab. This research experience can also prove to be helpful if you decide to move to the industry after your PhD. But the industry, especially startups, is also not a bad place to start your career as it exposes you to rapid product development cycles.

Andrey: Try and do some research! It’s easy to think you’d like research theoretically, but the reality of how it gets done is likely different from what you’d imagine having not given it a try.

---

Someone who’s decided to go to grad school but isn’t there yet?

Derek: The choice of advisor and lab may influence the next 4-6 years of your life more than any other choice.

Kiran: Choose a research topic and a research lab where you connect the most with people, especially the supervisors and your future colleagues.

Andrey: Don’t worry too much about knowing your exact thesis topic early on. Ask for advice from lots of people. Decide on your advisor and lab carefully.