Hardik Meisheri has been a researcher at TCS Research, Mumbai, India, since 2016. 

Hardik is working on how to use deep reinforcement learning to solve problems in the real world, such as supply chain optimization. In addition, he has worked extensively on sentiment analysis over noisy text.

INDIAai interviewed Hardik to get his perspective on AI.

What inspired you to seek a career in AI as an Electronics and Communication Engineering student?

As a part of our final year projects Bachelor's, we developed a smart wheelchair from the ground up, which can be controlled by speech, joystick and neck movements. We used micro-controllers and sensors to achieve this along with Hidden Markov Model for speech recognition. The primary motive for this project was to provide an affordable solution for quadriplegic people. After graduating, this problem stayed with me, and I wanted to take it further, where just thoughts could control a wheelchair. Stephan hawking's chair partially inspired it. It prompted me to look for solutions in Learning theory where each wheelchair could learn the nuances in the thought signals and act accordingly. It led to me taking up a master's in Artificial Intelligence.

What were the initial difficulties you encountered during this transition?

Coming from an Electronics background, it took a lot of work to wrap my head around many terminologies in most research papers. In addition, I was used to programming in Embedded C and MatLab, which has altogether different paradigms than Python or R. Although I ended up coding most of my thesis work in MatLab, it could have been more efficient. On top of that, in 2014-15, no stable libraries were available today, such as TensorFlow and PyTorch, for coding neural networks. We used to work with Theano and Caffe, which were riddled with bugs. Many Traditional Machine learning algorithms are derived from statistics, and I had to spend quite a lot of time going back and forth over multiple books to understand them.

Tell us about your brain-computer interface research using electroencephalograms.

EEG signals obtained from users are highly noisy and exhibit inter and intra-user variance due to many factors (sensor noise, mixed thoughts, mood, temperature, etc.). Nevertheless, we can use these EEG signals to infer the motor imagery tasks and, in turn, used to control any device. Common Spatial Pattern (CSP) is a widely used algorithm to extract features from the Spatiotemporal nature of EEG signals (22 channels spanning the whole scalp). However, it is prone to noise and overfitting. I proposed regularized CSP based on the Frobenius norm and used Type-2 Neuro-Fuzzy Inference System for classification. It achieved much better results in two-class and multi-class classification across multiple users. This paper got the best paper award in 2016 ISCBI, Switzerland.

What is your day-to-day role as a researcher at TCS?

The primary expectation is to create value by planning and managing R&D programs towards business advantage through innovative offerings and visibility through publications. As a senior and team lead member, I simultaneously drive 3-4 projects. These projects are a mix of both fundamental research and applied research which might be oriented to solve any business problems.

Most of my time is spent coding and running experiments on projects I am part of. The rest of the time goes into meetings, brainstorming new ideas, reviewing codes, mentoring juniors, giving/presenting talks internally and externally and writing papers.

What qualities do you expect from a newcomer working in the AI field?

So I generally consider two roles in today's AI field. Researcher and Practitioner. For a researcher, two crucial aspects are self-motivated and Research Aptitude. Research Aptitude is willing to dig deeper into any code or algorithm or comb through papers till satisfactory answers are received.

For practitioners, the most important thing is to know what is happening behind the libraries they are using. Most often than not, newcomers would use any existing library which would work perfectly fine with the toy dataset. But while solving the real-world problem, you often need to tinker and change default parameters. Practitioners should know the effect of those parameters on the algorithm and not treat those algorithms as black boxes.

Tell us about your AI research objectives for the future.

Language has facilitated interaction among humans, which has proved to be one of the most fascinating and significant steps in achieving Artificial General Intelligence (AGI). Recently, there has been rapid progress in developing general-purpose large language models (LMM) in natural language processing (NLP), which has led to a similar milestone as “Imagenet” in computer vision for NLP. Furthermore, recent LMMs such as GPT-3 and Galactica provide fascinating insights into learning capability using language alone. My research goal is to work on the intersection of RL and NLP. It can be crucial in developing a generalized framework for transfer and multi-task learning and pave the way towards AGI.

What advice would you give to students and professionals who want to work in artificial intelligence?

The AI field is vast in terms of breadth to which it is applied, such as NLP, Vision, RL etc. All of these will have different nuances and tricks. However, all of them would have a basis in foundational Machine learning. Therefore, I suggest people go through and understand Probability, Linear algebra, Statistical learning theory and the basics of calculus and convex optimization.

Can you recommend AI books or research publications to people just starting in the field?

Books:

1. Mathematics of Machine learning by Marc Peter Deisenroth, A. Aldo Faisal, and Cheng Soon Ong
2. Pattern Recognition and Machine Learning by Christopher M. Bishop
3. Reinforcement Learning an Introduction by Richard Sutton and Andrew Barto
4. Deep Learning by Ian Goodfellow and Yoshua Bengio, and Aaron Courville

Some inspiring seminal papers

1. "Attention is all you need" Vaswani et. al
2. "Mastering the game of Go with deep neural networks and tree search" Silver et. al
3. "Long Short-Term Memory" Hochreiter et. al
4. "On the Measure of Intelligence" Chollet et. al