Postdoctoral Researcher Soumyabrata Pal works at Google Research in India. At the University of Massachusetts Amherst’s Department of Computer Science, he earned his PhD.
He interned as a research assistant at the Ernst & Young AI Lab in Palo Alto and as an applied scientist intern at Amazon Search in Berkeley in the spring of 2020.
To learn more about Soumyabrata’s viewpoint on AI, INDIAai spoke with him.
You studied Electronics and Electrical Communication Engineering; how did you end up in AI?
During my undergraduate studies at IIT Kharagpur, students had the option of selecting extra or optional courses from different disciplines. I randomly enrolled in the CS department’s Intro to Machine Learning course and enjoyed the instructor’s approach to teaching statistics and algebra. I continued by enrolling in the advanced NLP and ML courses that were available, and in the end, I was able to make arrangements for my final year B.Tech Project to be in the CS department. During my college, it led me along the route of ML Theory.
However, a little-known but crucial course in EE is required for the theoretical aspects of ML. Information Theory is a course that is a sibling of Statistics. I thoroughly enjoyed the course (with the support of a superb professor) and stated in my SOP that I want to become an information theorist. Fortunately, my PhD advisor is a renowned information theorist who specialises in machine learning.
What does a graduate research assistant typically do in a day?
In my role as a research assistant, a typical day typically comprised reading an intriguing article or book or attempting to figure out a proof for a conclusion I had in mind. The theoretical guarantees in my work required to be supported by some small-scale experiments, so occasionally I also had to code those. It differs slightly from applied ML/NLP professionals, who spend the most of their day writing up various algorithms.
What difficulties did you first face as a graduate research assistant? Exactly how did you manage them?
When I started as a Graduate Study Assistant, I had no prior experience conducting any sort of ML Theory research. It was as though I had been dropped into an ocean with the goal of teaching myself to swim. It was difficult at first, but my advisor helped me get going. It greatly aided me to read timeless, authoritative publications that are written from scratch. When I need to learn a specialised area of ML Theory, I still make an effort to read books first (instead of recent publications).
Tell us about your PhD’s area of study.
Theoretical machine learning and applied statistics are my main areas of interest in study. I’m interested in creating algorithms that have theoretical or logical assurances and can direct real-world solutions to pertinent issues.
Consider a high-quality DSLR camera’s raw image, which can have a size of more than 50 MB. However, there are other image compression techniques, such PNG and JPEG, that can drastically reduce the size of the image without altering how it appears to the human eye. Thus, we can infer that a significant portion of the raw image’s pixels are useless, and it would be beneficial to create algorithms that can effectively compress and restore such images. Another illustration would be streaming services like Netflix, which provide millions of subscribers with individualised recommendations. Because many people have comparable tastes, collaborative filtering can combine their scores to get superior suggestions. However, the similarities and preferences must be promptly learned online because they are unknown.
The goal of my research is to create learning algorithms. Since speech and image signals are sparse, for example, these data structures either naturally arise in many applications or are frequently incorporated into data.
What progress has India made in ML and AI? What do you consider to be our strengths and shortcomings in this situation?
I think India is currently performing incredibly well and is catching up to western countries. India’s strength is that there are top-notch researchers working on ML in numerous institutes around the country. However, based on what I gather, India has a significant gap between theory and practise. The business community and colleges need to collaborate more in order to close this gap, which is not currently happening. In contrast, Silicon Valley in the US, which is renowned for its innovative capabilities, has regular collaborations between two of the greatest institutions in the world and business professionals.
What should Indian universities improve, in your opinion?
In my perspective, Indian institutions need to develop in two areas: 1) Increase collaboration with industry and comprehend practical, relevant challenges 2) Develop a more developed postdoc culture in India, which I believe is lacking in comparison to the west.
How did you respond to difficult circumstances as a researcher, such as paper rejection? How do you keep your composure?
It’s difficult to deal with rejection, and I still believe I could do better. It is still a possibility to revise the article and resubmit it. A better article version will have a greater impact because we aim to advance research.
What guidance do you have for those looking to work in the field of artificial intelligence research? What should they concentrate on to advance?
These days, there are a lot of prospects for AI research, both in academia and industry. I suggest taking ML/AI courses and starting with the fundamentals. Modern computing has made it simple to train sophisticated models and use them in real-world situations. But developing intuition and understanding must come first. Anyone who wants to work in the industry can benefit from the enormous selection of online courses that are available.
What books and academic articles have had a big impact on your life?
Three books had a significant impact on how I understood machine learning. They are listed below:
Gareth James, Daniela Witten, Trevor Hastie, and Ryan Tibshirani’s introduction to statistical learning, deep learning by Aaron Courville, Ian Goodfellow, and Yoshua Bengio
By Noga Alon and Joel Spencer, The Probabilistic Method.
