Machine learning researcher as a high-energy physics masters graduate [closed]

Question

From what I understand, the majority of novel machine learning (ML) research is in applications such as natural language processing (NLP) and computer vision (CV). How can I focus my career on becoming a researcher in these areas of ML?

My current background is in ML applications in high-energy physics, having completed my bachelor's and master's theses on this topic, along with a few internships with the CERN LHC experiments. To be most effective in ML research, would it make sense to obtain a masters in computer science, or could I use my time better given my current background?

I am very interested in pursuing a PhD with a focus on ML in high-energy physics. Would this be the optimal path to develop my career as an ML researcher, or would there be a better choice given my background?

One concern is that ML developments are relatively slow in high-energy physics, and I might not be making the best use of my time. Additionally, I am not sure if the skills I obtain will be the best match for becoming an ML researcher compared to a purely ML-focused PhD. However, I have a lot of knowledge in this domain and could contribute better through my insights compared to other scientific fields. How useful is the domain knowledge I have in developing my ML research skills? Do I benefit from being in a specialised domain as compared to more general domains like NLP or CV? Could I even pursue a purely ML-focused PhD with my background?

I have also considered the path of joining a tech company and working my way up to a research position by reading papers, implementing them in solutions, and gaining research skills as I do this. How would this path compare to doing a PhD? Would I be in a good position to effectively develop the skills I need for ML research?

To summarise, I see the options as follows (not in any particular order):

• Pursue a PhD in high-energy physics focused on ML
• Seek a purely ML-focused PhD
• Gain computer science specialization (Master's degree), then pursue a PhD (either purely ML or ML in high-energy physics)
• Join a tech company developing state-of-the-art ML in the company's domain

What would be the most optimal option to become an ML researcher? What are the potential pros and cons of each option (effectiveness, risk)? Is there another good option that I am not considering?

Answer 1

What would be the most optimal option to become an ML researcher?

It very much depends what sort of an "ML researcher" you want to be. Being a professor who studies the theoretical underpinnings of ML is very different from being an engineer who tweaks ML algorithms to improve a gizmo's performance. So to find the optimal path, you may need to constrain the objective.

On the other hand, if your objective is not so constrained, you don't necessarily need an optimal path. When people tell us they want to be a professor of physics, our advice is usually: "it's a very hard path, there are only a few spots each year; you need to do everything perfectly and also get lucky." But ML research positions are relatively easy to find right now (at least compared to physics professorships), so if you follow your interests and do excellent work, you will likely have an path forward.

To your options:

Pursue a PhD in high-energy physics focused on ML

This is what I did (though this was mostly before the ML craze). Was it the most efficient path? Probably not, but I have no regrets. Some pros:

• In high-energy physics specifically, most of the work involves scientific computing / writing code. If you take the time to really learn this stuff, it is an extremely solid background. I've never taken a programming course in my life, but my software and Linux skills after graduation were super strong, even compared to "real" software engineers.
• Learning about physics and math is very useful for certain application domains. When the subject matter experts start talking about partial differential equations and Bragg scattering, I am still at home. It's much easier to learn ML "on the job" than to learn physics and math "on the job."
• Personally, I found getting a PhD to physics to be very rewarding. I wouldn't want to go back to that life, but I'm glad I did it once; no unanswered questions or unfulfilled dreams.

Seek a purely ML-focused PhD

This is obviously the most direct path. I would ask yourself: where will you do the most exceptional work?

Gain computer science specialization (Master's degree), then pursue a PhD (either purely ML or ML in high-energy physics)

I see no reason to get another master's, unless you need it to get admitted to a good PhD program.

Join a tech company developing state-of-the-art ML in the company's domain

This is another direct path, though it has a few distinct considerations:

• Do you want a PhD? You probably don't need one; a few of the smartest researchers I know never got their PhD. But if you want one, it's probably easier to get it now; it's unlikely that you'll take time off later to go back to school, and trying to get a PhD while working is very difficult.
• Can you get a researcher job with only an MS? Such jobs are out there (I'm one of the interviewers!), but they are rare. And if you accept a different position, the odds that you successfully transfer into an ML research position are not super high.
• Do you want to keep the door open to a professorship? This will probably close the door, unless you come up with amazing breakthroughs that make you famous.
• Do you like money? This will pay much better, and not having to spend years earning a graduate student stipend will greatly improve your lifetime earnings.
• What kind of work do you want to do? There are some "theoretical" positions out there even in industry, where you might be doing theorem/proof or calculating theoretical guarantees. But as you can imagine, most positions are more applied: how we can use ML to solve [some problem]. Some of this is great, and you end up working on very difficult, interesting problems. But it's important to choose carefully: many positions are just engineering work, not really research, and some of it might be researchey but not in a very fulfilling area (e.g., advertising, insurance).