Soumyabrata Pal is a Postdoctoral Researcher at Google Research, India. He completed his PhD in the Computer Science Department at the University of Massachusetts Amherst.

He was a research intern at Ernst & Young AI Lab in Palo Alto and in Spring 2020 as an Applied Scientist Intern at Amazon Search, Berkeley.

INDIAai interviewed Soumyabrata to get his perspective on AI.

How did you get into AI as an Electronics and Electrical Communication Engineering student?

At IIT Kharagpur (during my undergraduate), students were given the luxury of choosing additional or optional courses from other departments. I took an Intro to Machine Learning course (offered in the CS department) on a whim and liked the teaching brand of mathematics/statistics. I followed up by taking advanced classes on NLP and ML that were on offer, and finally, I was also able to arrange for my final year B.Tech Project to be in the CS department. It brought me on the path of ML Theory during my undergrad.

Still, one course offered in EE isn't very well known but is very important for the theoretical parts of ML. The course is named "Information Theory" and is a brother to "Statistics". I loved the course (helped by an outstanding professor) and wrote that I want to be an Information Theorist in my SOP. It was fortunate that my PhD advisor is an esteemed information theorist working on ML.

What is a typical day in the life of a Graduate Research Assistant?

A typical day in my life as a Research Assistant usually involved reading an exciting paper/book or trying to work out a proof for a result that I had in mind. However, on some days, I also needed to code to do some small-scale experiments to support the theoretical guarantees in my work. It is slightly different from applied folks in ML/NLP, who spend most of their day coding up different algorithms.

What were your initial challenges as a Graduate Research Assistant? How did you handle them?

I did not have any research experience at all in ML Theory when I joined as a Graduate Research Assistant. So, it was as if I was thrown into an ocean with the expectation that I would learn to swim. Initially, it was challenging, but my advisor helped me ramp up. Reading classic relevant books that start from scratch helped me a lot. I still try to read books first (instead of recent papers) when I need to understand a niche field in ML Theory.

Tell us about your PhD research area

My primary research interests are Theoretical Machine Learning, Applied Statistics. I am interested in designing algorithms with theoretical/provable guarantees that can guide practical solutions to relevant problems.

For example, consider a raw image captured by a high-quality DSLR camera whose size can easily exceed 50 MB. However, many image compression schemes such as PNG and JPEG are available that can reduce the image size significantly without making any difference to the human eye. Hence, we can conclude that many pixels in the raw image are irrelevant, and it would be helpful to design algorithms that can compress and recover such images provably. As another example, consider streaming engines such as Netflix, which makes personalized recommendations to millions of users. Many users have similar tastes, so their ratings can be aggregated to make better recommendations - collaborative filtering. However, the similarities and preferences are unknown and need to be learnt quickly online.

My research is focused on designing learning algorithms. Such structures in data either occur naturally in many applications (for instance, image and speech signals are sparse) or are often embedded into data.

How is India doing in the areas of AI and ML? In this situation, what do you think are our strengths and weaknesses?

I believe India is doing exceptionally well nowadays and is catching up to western nations. India's strength lies in the fact that there are world-class researchers in many different institutions who are working on ML. However, from my understanding, the main weakness is a huge gap between theory and practice in India. To close this gap, the business world and universities need to work together more, which is not happening now. In contrast, Silicon Valley in the US, famously known for its innovation capabilities, has two of the best universities in the world regularly collaborating with industry practitioners. 

What aspects do you believe Indian universities should improve?

Two aspects need to be improved in Indian universities, in my opinion, 1) Increase collaboration with industry and understand practical, relevant problems 2) Improve the postdoc culture in India, which I feel is not as mature as in the western countries.

How did you handle challenging situations like article rejection as a researcher? How do you compose yourself?

Rejection is tough, and I still feel I am not good at handling them. Still, it's a chance to make changes to the article and try again. Since we want to improve science, a better article version will have more effect. 

What advice do you give individuals who wish to work in artificial intelligence research? What should they focus on to progress?

There are many opportunities in AI research nowadays, be it in industry or academia. I advise taking courses on ML/AI and understanding the basics first. Nowadays, with the advent of modern computing, it has become easy to train complex models and apply them. First, however, it is essential to develop intuition and understanding. The mammoth number of available online courses can help anybody who wishes to enter the field.

What research papers and books have significantly impacted your life?

There are three books which were impactful for me in understanding ML. I have listed them below:

1. An introduction to Statistical Learning by Gareth James, Daniela Witten, Trevor Hastie and Ryan Tibshirani
2. Deep Learning by Aaron Courville, Ian Goodfellow, and Yoshua Bengio
3. The Probabilistic Method by Joel Spencer and Noga Alon.

Apart from these, there are websites. 

(Machine learning mastery, for example) and blogs that are sometimes super useful for quickly gaining a high-level understanding of some complicated topics.