Microsoft Research India
AsianScientist (Nov. 26, 2020) – Imagine this: what if you suddenly woke up with the superpower to answer questions with billions of choices in milliseconds? Would you try to see who among the world’s population should first receive the potential COVID-19 vaccine? Would you try to forecast which grocery or retail items you’d likely need even before you’ve checked your room and cabinet? Or would you want to predict the next words uttered by your favorite public figure?
For more than half a century, scientists have been trying to build machine learning algorithms to answer such multiple-choice questions with a high level of uncertainty. Unfortunately, despite decades of progress, state-of-the-art algorithms could only pick the correct subset of answers for questions with just thousands of choices. Enter Dr. Manik Varma, a Partner Researcher at Microsoft Research India.
In 2013, Varma and his colleagues developed an algorithm that could answer queries with millions of choices—kickstarting a new field of machine learning known as extreme classification in the process. The secret lies in quickly eliminating the vast majority of irrelevant options, allowing the algorithm to focus on the few hundred remaining choices. Among numerous other applications, extreme classification has been used to generate increasingly sophisticated web searches, more tailored advertising and personalized recommendations.
Since then, Varma has gone even bigger. His team’s award-winning Slice algorithm, for instance, can process problems with 100 million choices in mere milliseconds. For his pioneering contributions to engineering, Varma received the 2019 Shanti Swarup Bhatnagar Prize for Science and Technology, an annual award given in India for notable science research. In this interview with Asian Scientist Magazine, Varma recounts his early beginnings in the field of machine learning and reveals his hopes for his research in the years to come.
- How would you summarize your research in a tweet?
I work on projects at the extremes of machine learning—ranging from deploying learning algorithms on chips smaller than a grain of rice to answering multiple choice questions with millions or billions of choices in milliseconds.
- Describe the research project that you are proudest of.
We started the area of extreme classification which studies multiple choice questions with an extremely large number of choices. It brought in new research problems that the classification community hadn’t been thinking about traditionally. It has found application in diverse areas including computer vision, natural language processing, information retrieval, etc. and has opened a new paradigm for key industrial applications in web search, computational advertising and recommender systems. Today, extreme classifiers are making billions of predictions daily, are helping millions of people be more efficient and are generating millions of additional dollars of revenue for many businesses worldwide.
- What do you hope your research will accomplish in the next decade?
I hope that our algorithms will one day reach a level where the product is delivered to your doorstep just as you realize you need it. Our algorithms should anticipate your requirements and provide you the necessary information at just the right time
- Who (or what) motivated you to go into your field of study?
I’ve been very fortunate to have found exceptional mentors in my family (my parents and uncle), my teachers (Andrew Zisserman at The University of Oxford and David Forsyth & Jitendra Malik at University of California at Berkeley) and my managers at Microsoft (P. Anandan and Sriram Rajamani).
My field of study has shifted a lot over the years. It was Anandan who encouraged me to switch from computer vision to machine learning more than a decade ago. For the last five years, Sriram has been mentoring and supporting me and encouraging me to try out applications in diverse areas. Recently, my students and collaborators have motivated me to keep exploring new fields, learning new things and building solutions that benefit people.
- What is the biggest adversity that you experienced in your research?
All sources of adversity have turned out to be opportunities in the long run. Having a limited budget forced us to improve our algorithms so that they could run on a single core of a standard machine rather than on huge clusters, and not having all the skills in my team sparked hugely beneficial collaborations with other teams.
- What are the biggest challenges facing the academic research community today, and how can we fix them?
Some of these challenges include making research relevant to society, communicating it to the layperson, ensuring quality, encouraging students to pursue their dreams rather than going after the highest paying jobs, strengthening undergraduate pipelines, encouraging researchers to take risks, modifying the system to allow for radically novel ideas to propagate faster, tackling integrity, fairness, ethics and inclusion, etc. Experts would know how to tackle these challenges better than me, but I suspect years of experimentation might be required.
- If you had not become a scientist, what would you have become instead?
I often fantasize about becoming a grand master at chicken chess but, realistically speaking, I don’t think I would ever have amounted to much.
- What do you do outside of work to relax? Do you have any interests and hobbies?
I used to love reading but had to give it up as I lost my eyesight. However, audiobooks have opened up a whole new world for me. I also like watching plays, going on history walks, discovering new food, swimming with my kids, travelling with my wife and learning about literature from my parents. I have also been training to be a Dungeons and Dragons’ dungeon master during the COVID lockdown.
- If you had the power and resources to eradicate any world problem using your research, which one would you solve?
I’d love to see a world where everyone can instantly get accurate, verified and reliable information in a form that is accessible to them while respecting security and privacy concerns.
- What advice would you give to aspiring researchers in Asia?
Something that has worked well for me has been to spend a significant amount of time identifying the right problem to solve. I am often tempted to solve problems in my comfort zone. However, I have found it more fun and rewarding to tackle impactful rather than easy problems.
In Asia, particularly, we have access to many problems which, if solved, can benefit millions of people or save millions of lives or generate millions of dollars in revenue or even put a person on Mars. However, challenging problems like these are also complex and risky. So, I try to have a long time horizon, try to put together a team with the necessary skills and structure the project to mitigate risk. Instead of stopping at the publication stage, I try to solve the problem end-to-end and go through multiple rounds of deploying the solution in the real world to maximize its benefit to people.
I am often afraid of failure but have come to realize that I have learnt more from my mistakes than from my successes.
This article is from a monthly series called Asia’s Rising Scientists. Click here to read other articles in the series.
Copyright: Asian Scientist Magazine; Photo: Manik Varma.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.