Meet the Researcher: Gaurang Gavai Discusses the Latest Advances in Machine Learning

Our own Aki Ohashi, Director of Business Development, recently interviewed PARC Researcher Gaurang Gavai to learn more about his interests and work in machine learning.

Gaurang, can you tell me a little about your background?

Sure thing. I’m originally from Bombay, but we moved around a lot as a kid. I spent a lot of time in Minneapolis, but also in Paris and Tokyo, and then back to Bombay. Ten years ago, I moved to Atlanta for graduate school, then spent some time in Chicago, and now in San Francisco. I feel like this is home for me.

Do you think your diverse background helps you in some way?

Definitely, I think it makes me a bit of a chameleon. I also notice this with some of my other friends in the Bay Area that have come from different places. I feel it’s easier for us to fit into new situations.

Can you talk about your education?

I obtained my master’s degree in machine learning at Georgia Tech, focusing on the theory of machine learning. My thesis was on active learning which is the idea of having humans in-the-loop with systems. Instead of training with large complete datasets, we often get data with very few labels or spotty information. In these cases, you can add a human into the loop and annotate only the most important data. It was actually really good timing that this area became immensely popular right as I was about to graduate.

Why did you pick active learning?

I was always interested in artificial intelligence. I did a little of that during my undergraduate years in Bombay, and my thesis was about finding new ways to do optical character recognition. It’s basically a solved problem today, but back then, people used to try to match templates and do interesting things with local binary patterns. Which is why I thought this new methodology of not having to hard-code all of these features and tell the computer what to look for was fascinating. I thought it was really an interesting idea to give the computer the opportunity to learn for itself. 

Now, active learning has progressed from having a human in-the-loop to having a neural network in-the-loop. In this meta-learning concept, instead of teaching it how to learn, it’s teaching it how to learn to learn. 

Would GANs (Generative Adversarial Networks) be an example of that?

Yes, but actually, all deep networks are like that. Let’s take facial recognition as an example. People used to compute the distance between the eyebrows and nose to use as a feature. But now, with facial recognition, you feed the images in, and the neural network figures out what features are important to identify people. The computer figures out new ways to infer things, which I think is really cool.

Where do you think this is going, say in 3-5 years?

That’s an interesting question. I see multiple routes. One, which PARC is also focusing on is, “how do you do real things with this?” Things like Facebook’s facial recognition and Google Maps work really well, but that’s because they have poured a lot of money into them. How can we get to those places much faster? I think this is an area where PARC really excels. We can take new situations or sparse data, and try to create real value from them. 

The second route is to understand and explain these models. Explainable AI (XAI) is an example of this. In certain subject areas, you can’t have a model that is opaque. You need to be able to understand what it is doing and explain very clearly how its decisions are made.

I think the third route is what society is obsessed with, which is complete automation. How do you move towards systems that are fully automated? That’s the most open-ended. People talk about the “singularity” in 2050, though I don’t know if that’s true or not.

Would the second and third routes work against each other?

I think it’s a very funny relationship. The example that comes to mind is martial arts, where we always talk about “strength” versus “balance.” I think those two things need to work together, though they pull from opposite ends of the spectrum. You need to be able to do both.

How did you end up at PARC?

Actually, I got a cold call from a PARC researcher on my way back from another interview. I didn’t know him, but he must have looked me up through my professor. He asked if he could ask a few questions, and I said, “yes,” thinking they would be a few screener questions from HR. It turns out, he dove right into a technical interview. I ended up doing two more phone interviews, came to PARC to do a talk, and then completed some more in-person interviews. That was a long day, but it was really good. On my way back to Atlanta, even with all of the offers I had, I knew PARC was the place for me.

Why is that?

There are two reasons. First, at PARC, there is a sense of academic freedom, which everybody in the building appreciates. Yes, we work in industry, yes, it’s applied, but you get to set research agenda and help shape the direction of things. The second is that we’re not working in an academic bubble. We are client facing, addressing real-world problems. It’s a really nice balance between academia and industry.

I have so many friends that also graduated at the same time. They’ve mostly all either gone back into pure Ph.D. academia or into industry like Facebook or Google. And it seems that they each lack one of those two things. This place is rare.

PARC Researcher Johan de Kleer once said, “There are two types of people that are here. You either come and leave quickly, or stay for a long time.” 

I understand. Though I’m not on the technical side, I feel the same thing, and thus, I’ve been here 11 years. Do you feel that you need to get a Ph.D.?

No, I don’t think so. The kind of experience and education that I am getting here is the kind that I want. I don’t want to super-specialize into a particular area. I’m a dabbler by nature, and do a little bit of a lot of things. Fortunately, you get to do that here. You can change projects every six months and you can spend time interfacing directly with clients. More recently, I’ve gotten more interested in working with clients, and that has almost immediately been reflected in my job. (Note: Gaurang traveled with Aki to Japan last fall to help run a workshop with one of our clients.) I like that you can “manifest your own destiny.”

How can a non-technical person learn more about AI?

There are two ways to learn: learning to understand and learning by doing.

For “learning to understand,” Coursera is great. I think Andrew Ng’s Stanford course is the best out there. I remember doing it in graduate school thinking that it was a great course, but I recently went back to it, and it is still very relevant, provides excellent teaching and is great in terms of theoretical understanding.

As for “doing,” it’s best to just jump in and do it. In graduate school, I used Kaggle, which I think is still very popular. It’s a data science platform that has competitions. They had one with NASA that had a $50K prize. You can form teams or do it alone. The nice thing is that they provide a lot of resources to get started. All of the data is provided. At the end of the competition, you can see the algorithms of the top teams.

What do you do outside of work?

I spend 4-5 nights a week in a form of martial arts similar to Jeet Kune Do. Part of the program is about giving back, so I also teach and help other students. I also play bass, guitar, and sometimes percussion in a blues band with a few of my friends. We play small house parties and such. I also like to travel as much as possible.

Where do you like to travel?

The next places I’d like to explore are Morocco and Egypt. The more out of the way, the better.

That’s great. I’ve not been to Morocco but Egypt was fabulous. I would also recommend Africa. In any case, is there anything else you’d like to tell our readers?

I know that people say this a lot, but I’ll just be blunt about it. I think PARC is really special. What we do well is work on hard problems that other places are not able to do. We’re really good at being on the cutting edge!

Why do you think that is? PARC is really small, about 200 people. Why are we able to do that?

I forget who said this but the phrase that’s stuck in my mind is, “PARC’s strength comes from the homogeneity in heterogeneity.” We all have a similar mindset about things, and similar things that we want to achieve, but we have such diverse backgrounds (physics, biology, computer science, pure math, social science) and we all come together to work on hard problems. 

Additionally, there is such little politics and hierarchy here. It makes it easier to get stuff done.

Thank you for taking the time to chat with me. Would you mind if people contacted you?

Absolutely. They can shoot me an email at ggavai@parc.com. I’d love to hear from people, learn about what they are working on, etc.