DataHack Radio #17: Reinforcement Learning with Professor Balaraman Ravindran
Reinforcement learning is an intriguing but complex topic to get your head around. It’s also one of the most promising skills a data scientist can add to their portfolio. Reinforcement Learning has sprung up some of the biggest ground-breaking developments in the last few years, including powering Google DeepMind’s popular AlphaGo program.
Who better to demystify the aura around this vast field than India’s foremost researcher on Reinforcement Learning? Yes, we’re talking about none other than the eminent Professor Balaraman Ravindran!
Professor B. Ravindran has an incredibly rich background in academic research, headlined by his work in reinforcement learning. His Google Scholar profile shows his research papers have been cited over 2,200+ times!
He has a penchant of explaining the most complex topics in words even beginners are able to grasp. It’s just one of the many reasons his talk at DataHack Summit 2018 was a super hit among our community.
In this podcast, Kunal speaks to Professor Ravi about his background, his interest and research in reinforcement learning, and the intricacies and nuances of this field.
All DataHack Radio episodes are available on the below podcast platforms. Subscribe today!
Professor Balaraman Ravindran’s Background
Neural networks were all the rage back in the 1990’s. Given Professor Ravi’s passion for computational models, it was inevitable he would start delving into this subject.
As he explored state-of-the-art neural networks during his Master’s (from IISc), he realized that the approaches were moving away from explaining how humans learn. That led to his foray into the world of reinforcement learning (and we are all really grateful for that!).
He started reading research papers on neuroscience as his fascination with reinforcement learning continued to grow. So what was the state of RL back then? This quote from Professor Ravi sums it up perfectly:
“I was the only person in my Master’s class who was working on Reinforcement Learning.”
One of the biggest drawbacks in India was that there wasn’t a single resource available to learn reinforcement learning. So Professor Ravi along with other researchers wrote a survey paper and emailed it out. This survey picked up pace in the community and even caught the eye of the Oxford press. They asked Professor Ravi and team to write a chapter on RL for their handbook on neural computation (published in 1996).
Ph.D in Reinforcement Learning and Working with Andrew Barto
This survey served as Professor Ravi’s entree into his Ph.D, which he successfully completed from the University of Massachusetts. His Ph.D advisor? None other than the great Andrew Barto!
One of the questions Professor Ravi and Andrew Barto wrestled with concerned the human psyche. How are humans so good at learning one problem and transferring it to another task so quickly? The duo attempted to solve this question for tasks that were similar in nature.
If you’re wondering what “similar” here means, you aren’t the only one! Professor Ravi’s research came to the point where he needed to formally define this word in the context of his research. That’s how they came up with the concept of abstract algebra and the mathematics of homomorphisms. A lot of transfer learning frameworks have actually been built on homomorphisms.
Here is his formal definition of similarity:
“Two things are similar if, for everything I can do for situation A, there is a decision in situation B that has similar effects.”
Professor Ravi was kind enough to share his Doctoral dissertation presentation on this topic which you can download here. It is a MUST-READ for anyone pursuing this field.
Later, while working with one his students, Professor Ravi discovered that this notion of similarity in RL was exactly what computer scientists call graphs. A fascinating insight into how the approach works!
“The mathematical notions of similarity that we define are all on an abstract model of the world.”
Research at IIT-Madras
Professor Ravi joined IIT-Madras’ faculty following his Ph.D. There he explored multiple areas of research since reinforcement learning had still not picked up steam. He explored domains like Natural Language Understanding and learning on graphs.
Circling back to RL, Professor Ravi has continued to work on homomorphisms at IIT-Madras. Another stream he has been working on is learning complex policies. The question he has been exploring deals with how to design agents that can mimic human thinking (the same question he was pursuing with Andrew Barto).
Hierarchical reinforcement learning frameworks have been another area of interest and research (including a ton of work on attention modeling). Complementing that is deep reinforcement learning (on ATARI games) which really took off in 2014. This gave Professor Ravi more complex domains to work on. He has explained the Deep RL concept using a superb analogy. This section is a MUST-LISTEN for everyone in data science. You will be able to understand the explanation even if you’re a relative beginner in this field.
This section is a nice microcosm of the magic behind Professor B. Ravindran’s teaching methods.
Another direction of research Professor Ravi has been looking at is going beyond rewards. We usually work with rewards when experimenting with a RL approach, right? But there are plenty of other signals apart from this in real-world scenarios. It’s critical to understand this if we are to integrate reinforcement learning in a human-centric world.
Professor Ravi’s Work in the Industry
What makes Professor B. Ravindran’s experience unique is that his work isn’t limited to academia and research. He has worked on multiple industry projects, including a few on:
- Language (NLU)
- Multi-modal learning
- Learning with networks
What about RL specific projects? Those do come along, but not as often as one might think. He has worked on RL projects where the optimization component is rolled into the problem (where there’s no pre-defined model).
Robert Bosch Centre for Data Science and Artificial Intelligence
Professor Ravi is also the head of the Robert Bosch Centre for Data Science and Artificial Intelligence at IIT-Madras. This was founded in 2017 with a vision to become an internationally renowned centre for data science research, where long-standing fundamental research problems, cutting across disciplines, are targeted and solved.
I highly recommend following their GitHub page to check out their latest work.
I had the pleasure of meeting Professor Ravi at DataHack Summit 2018. He is a very down-to-earth person with an incredible enthusiasm for this field. He has an infectious joy about him and hearing him talk about RL feels like a dream come true.
All these qualities come out in this episode as well. What an awesome episode with one of the top people in our community. Happy listening and do share your feedback with us below.