About Me

I am CS PhD student in robotics & machine learning at the Robotics & Embedded Systems Lab (RESL) , advised by  Gaurav Sukhatme . ...

Entering an ML-based PhD from industry

In recent times, people have asked me about PhD programs, especially students from Georgia Tech OMSCS. A number of posts online exist about applying for PhD programs and the breakdown of logistics and costs involved; many are directed towards people going from undergrad or MS directly. But only a few websites talk about entering PhD from industry. Moreover, in topics like machine learning and AI, there are so many people applying that you have a small chance of getting in to a good program unless you have some prior research experience. Lack of this experience is disadvantageous for folks coming in from industry (although having industrial experience is definitely helpful). If you are in that category, this post talks about what you can do to improve your profile for graduate school. I haven't tried all of them, but it is a collection of things I have read and heard from people.

Research Projects with a local university:
Try to do research projects in the domain of your interest with a local university. You might be able to look up lab webpages and then contact students or post-docs who might be able to talk to you about potential research projects. You will get a glimpse of what research looks like, see how academia works, and maybe even get a recommendation from the professor whose lab you worked for.

Code Samples:
It's definitely good to have code samples that are publicly viewable to the admissions committee. Your work in industry may not be publicly accessible, but you should do online courses and side projects that help your profile. This would also help you for job applications in industry.

Reproducibility of Papers:
If you are lacking research experience or relevant publications in ML/AI, one thing that can really help is reproducing the results of well-known papers. Try to get the same results of these papers on your machine, and put your reproduction code on github. Add your own take on implementation details, etc. There is also an Machine Learning reproducibility challenge in conferences, to which you can submit your reproductions and get a quasi-publication. This will help your github profile, mentioned above. Maybe you'll even get to go to the conference and meet professors in person!

Even if your work above isn't accepted, you should definitely go to conferences and meet academicians. You have to go out on a limb and introduce yourself, which can be hard at times, but it's definitely better than emailing them cold. It helps your application when you've met professors whom you want to work with. Even if you don't eventually work with them, they might even be generally good advisors or mentors. I met a professor who taught online ML and RL courses, who really helped me think through grad school applications.

These things will make your application stand out, and give the admissions committee a good impression of your intent to do research. Add industrial experience in your field to this mix, and you have a solid chance of getting into the labs you want!

This post was inspired by this page (google translated English version).