Before joining TCS, Mayank Baranwal was a postdoctoral scholar in the Department of Electrical and Computer Engineering at the University of Michigan, Ann Arbor. 

IIT Kanpur awarded him the Institute Silver Medal in 2011, the ME Outstanding Publication Award by the University of Illinois in 2017, and the Young Scientist Award by Tata Consultancy Services in 2022.

INDIAai interviewed Mayank to get his perspective on AI.

How did a mechanical engineering graduate end up in artificial intelligence? How did it all begin?

The field of AI encompasses various topics and applications, including automation, improved perception, learning, and cognition. It includes disciplines such as game theory, optimization, and control theory and can be applied to various fields, including mechanical engineering. Many people may have a narrow view of AI, mistakenly equating it solely with deep learning. However, AI encompasses a much broader range of technologies and techniques that we can use to assist with decision-making, perception, learning, and cognition.

During my undergraduate studies at IITK, I became interested in mathematics and decided to pursue a graduate degree in a related field, precisely control theory and its applications. I took many mathematics, electrical engineering, and computer science courses, including fundamental topics such as graph theory, reinforcement learning, game theory, and optimization. While studying, I also became aware of the rapid advancement of deep learning. As a result, I became interested in exploring some of the fundamental problems in AI, including neural graph networks and their limitations. In addition, I used my background in control theory to develop more efficient and reliable optimization algorithms. During my postdoctoral work at the University of Michigan, I had the opportunity to work on a project related to using deep learning for accelerated drug discovery, which allowed me to apply my knowledge to real-world problems involving modern learning-based methodologies. Overall, while I have always been interested in fundamental issues in AI, in recent years, I have also focused on learning-based approaches.

Can you describe your work at NASA's Jet Propulsion Laboratory?

During my master's degree, I interned at JPL's Dynamics And Real-Time Simulation (DARTS) lab. There I was tasked with developing coordinate-free governing equations for systems with variable mass, such as rockets and balloons, and verifying their accuracy using NASA's high-fidelity simulator, DSENDS.

Around this time, NASA announced its Asteroid Redirect Mission, which aimed to visit a near-Earth asteroid, collect a boulder from its surface, and redirect it into orbit around the moon. Another project I worked on at JPL involved accurately modelling the rough, spiky surfaces of asteroids using a small number of parameters.

Tell us about your PhD research topic and results.

My research interests are multi-disciplinary, as demonstrated by my doctoral thesis, which covered two distinct topics:

(a) the use of entropy-based methods for combinatorial optimization and

(b) the development of a distributed architecture for robust and optimal control of microgrids.

Combinatorial optimization problems are generally NP-hard, and most existing heuristics could be faster or more suboptimal. During my doctoral studies, I developed a facility location viewpoint of several combinatorial optimization problems, such as the travelling salesman problem, vehicle routing problem, multiway k-cut problem, etc. We showed that Shannon-entropy-based ideas could result in efficient, high-quality suboptimal solutions to several discrete-optimization problems.

The second part of my doctoral work dealt with scalable, optimal and robust control of microgrids. The increasing use of renewable generation and distributed energy resources (DER), such as residential solar and home energy storage, and customers' changing energy usage patterns are leading to more significant uncertainty and variability in the electric grid. Therefore, a comprehensive approach is needed to address the challenges to the reliability and power quality of the electric grid caused by widespread renewable power generation. This approach should consider centralized cloud-based and distributed peer-to-peer networks and allow for the coordinated response of many local units to adjust energy consumption and generation, satisfy physical constraints, and provide ancillary services as needed. My doctoral work utilized concepts from nonlinear and robust control theory. I tested it across various scenarios using actual physical devices such as photovoltaics, battery storage inverters, and home appliances. A vital feature of the proposed architecture was its flexible plug-and-play architecture, which allowed for easy engagement and disengagement of devices and small power networks from other power networks or the grid.

As a Visiting Professor at the National Institute of Industrial Engineering and IIT Bombay. What drew you away from academics and towards the AI industry?

I have always retained my passion for academia. I still actively teach, interact with students and faculty, and conduct research in academic and industrial settings. I don't view the distinction between academic and industrial research as a separation between fundamental and applied research. On the contrary, TCS Research, where I currently work, is a place that fosters intellectual stimulation and the pursuit of problems that have the potential to have both short-term and long-term impacts, similar to Bell Labs. I am fortunate to enjoy the benefits of both academic and industrial research and continue teaching and working on open-ended problems.

What initial challenges did you experience during this transition? How did you overcome them?

The transition went smoothly, but the major challenge was the unforeseen COVID-19 pandemic. My transition from the US to India overlapped with the advent of the COVID-19 pandemic in the US. I couldn't find a flight to travel back home and was eventually repatriated on the Vande Bharat flight in June. I also had a few excellent academic jobs offers around then (along with my TCS offer), but the universities were also under strict lockdown. Fortunately, I got to know a few top TCS Research scientists and decided to try that. I was pleasantly surprised by the research environment at TCS. TCS welcomed original thinking and enabled me to work on challenging open-ended problems (quite unlike an industrial organization), too. It was enough to keep me engaged with the industry. Of course, I have a lot of friends and batchmates, who are full-time faculty at IITs and IISc, so I have always kept in touch with academia in general.

What do you do daily as a Senior Scientist, Data and Decision Sciences at TCS Research?

At TCS Research, I lead a team of about five researchers with varied interests in the broad fields of AI. I am leading multiple projects at TCS. On the applied side, our team is working on the following:

• Learning-based control of power networks.
• Prediction of term/preterm deliveries based on vaginal microbial profile.
• Large-scale multi-robot task allocation in warehouses.
• Scheduling of refinery operations.
• Inexpensive data-driven spatial transcriptomics.

On the foundational front, I'm leading an effort on the analysis and design of novel optimization algorithms using ideas from control theory. Besides, I engage with other teams on several different research threads.

As such, just as in academia, I mentor researchers, collaborate and brainstorm with them daily. We also often interact with business units that take our research outputs and produce them for business needs. Besides, there are administrative responsibilities, such as interviewing candidates, working on research proposals, assisting other teams, conducting internal workshops, etc., that one needs to be involved with.

What characteristics do you seek in a beginner in the field of AI?

It has to be an eagerness to understand and learn. And that is generally true with almost all other fields, too. One of my teammates, when he joined us, needed to gain a background in AI or deep learning. But with his sincerity and frequent engagements with us, he was able to pick up some of the most advanced concepts in AI. He is now the primary co-author of our work on learning-based control of power networks that will appear in this year's AAAI (top AI conference). When I hire candidates for my team, I look for their sincerity, ability to articulate some advanced topics and eagerness to learn.

What advice do you have for students and professionals considering a career in artificial intelligence?

My primary advice for them will be to avoid getting seduced by what everyone else is doing. Instead, they should work on creating their niche. Identifying and working on fundamental problems is much more rewarding. For example, I have interacted with folks who would boast of implementing a complicated transformer-based architecture for some NLP tasks. But I would need to learn basic concepts, such as what makes the loss surface non-convex, why deeper networks are better than wider networks, etc. It is super important to think deeply and spend a good deal of time learning the various nuances. 

Can you recommend any AI books or research papers to newcomers to the field?

The usual suspects include the Deep Learning Specialization by Andrew Ng or the Neural Networks and Deep Learning course by Geoffrey Hinton. While Andrew's lectures give an excellent summary of the basics and applied aspects of deep learning, Hinton's lectures are more thought-provoking.

As I said, not all AI is deep learning. One of my favourite books is "Probabilistic Graphical Models: Principles and Techniques" by Daphne Koller and Nir Friedman. Below is an incomplete list of books that are potentially beneficial for beginners in AI:

• "Probabilistic Graphical Models: Principles and Techniques" by Daphne Koller and Nir Friedman
• "Reinforcement Learning: An Introduction" by Richard Sutton and Andrew Barto
• “Neuro-Dynamic Programming” by Dimitri Bertsekas and John Tsitsiklis
• "Deep Learning" by Aaron Courville, Ian Goodfellow, and Yoshua Bengio 
• "Elements of Information Theory" by Joy Thomas and Thomas Cover

For the research papers, I am particularly fascinated by the seminal work on Graph Convolutional Networks by Kipf and Welling. GCN has now become a household name in the deep learning community. Besides, there is a set of related works on optimization that have captured my interest. Below is an incomplete list of research papers that have impacted me.

• Welling, Max, and Thomas N. Kipf. "Semi-supervised classification with graph convolutional networks." In J. International Conference on Learning Representations (ICLR 2017). 2016.
• Kingma, Diederik P., and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).
• Goodfellow, Ian, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. "Generative adversarial networks." Communications of the ACM 63, no. 11 (2020): 139-144.
• Wang, Yue, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. "Dynamic Graph CNN for Learning on Point Clouds." ACM Transactions On Graphics (tog) 38, no. 5 (2019): 1-12.
• Wibisono, Andre, Ashia C. Wilson, and Michael I. Jordan. "A variational perspective on accelerated methods in optimization." proceedings of the National Academy of Sciences 113, no. 47 (2016): E7351-E7358.
• Martens, James, and Ilya Sutskever. "Training deep and recurrent networks with hessian-free optimization." In Neural networks: Tricks of the trade, pp. 479-535. Springer, Berlin, Heidelberg, 2012.
• Chen, Ricky TQ, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. "Neural ordinary differential equations." Advances in neural information processing systems 31 (2018).
• Tishby, Naftali, Fernando C. Pereira, and William Bialek. "The information bottleneck method." arXiv preprint physics/0004057 (2000).