Vinayak Gupta is a research scientist at IBM. He completed his PhD at the Indian Institute of Technology (IIT) Delhi. 

Vinayak earned the best doctorate work award at AI-ML Systems 2021 and was a runner-up in the Nasscom AI Game-Changers of India 2022 competition.

INDIAai interviewed Vinayak to get his perspective on AI.

What inspired you to pursue a career in AI?

During my undergraduate years at IIIT Jabalpur, I was deeply interested in robotics and IoT. To highlight a few, I was a part of the college Robocon team, and my internship project involved designing a robotic hand for holding an ultrasonography probe. In addition, a research project involved working on wireless sensor protocols and how we can leverage AI models to overcome human-defined functions. Since all these techniques were relatively new, most of my work involved researching, i.e., studying the existing literature and identifying suitable techniques. So naturally, a PhD was the next progression to continue my career in research. At IIT Delhi, I was incredibly fortunate to be advised by Prof. Srikanta Bedathur, who gave me a detailed overview of the limitations and strengths of current deep-learning models. After a much deeper analysis, we realized they were significant drawbacks in the existing techniques for performing real-world applications such as data mining, retrieval, and forecasting. It further inspired me to pursue a career in AI research and designing models that can solve these high-impact problems.

Could you please highlight the key issues you are researching?

My ongoing research can be divided into two segments:

(i) continuing my PhD research of designing scalable models for time-series data; and

(ii) the ongoing projects at IBM Research.

Here, I provide a summary for each:

• Deep learning for time series: I am continuing my PhD research in designing scalable and accurate models that can solve various problems, including temporal sequences. These problems include ways to overcome missing events in temporal sequences, scalable sequence retrieval, and modelling the actions done by users in activity videos. Thus, most of this work involves understanding how the time series evolve, the influence of past events on the current timeline, and predicting and forecasting future events. 
• System for Data Practitioners: At IBM, I am working on a project that involves understanding the needs of data practitioners in natural language. In detail, here we tried to understand the type of problems that data practitioners such as data scientists and data managers can have and design ways to solve all their problems in natural language.

What were your initial challenges as a graduate student in AI research?

The major challenge was to understand the psychology of delayed gratification, i.e., the act of resisting an impulse action for an immediate reward in the hope of obtaining a more-valued reward in the future. Mastering this skill is necessary to publish high-quality papers. Scientific research is a lengthy process that includes identifying impactful problems, collaborating with others, writing an initial draft, and getting it reviewed at leading AI venues. These steps can take a significant amount of time to finish. As an undergrad, I expected results when I spent a task. Therefore, I had to overcome the tendency to be impatient and expect results the moment I finished a job.

We all say that machine learning and deep learning are new technologies that have changed the world. But we often don't notice the unique problems they bring with them. One of them is how these fast technologies affect the environment. What do you think about that?

I understand the concerns associated with technologies that necessitate energy-intensive computations and their environmental impact. Similarly, training ML models requires significant energy consumption and, thus, may have a large carbon footprint. For reference, a study showed that training an ML model, such as a Transformer, can have significantly higher CO2 emission than a car in its lifetime. Therefore, there has been a recent surge in models with significantly fewer trainable parameters. There has also been a surge in designing reusable models, i.e., the parameters can be used for different tasks by minor fine-tuning. I sincerely hope that these ideas will reduce the carbon footprint of training larger ML models.

Digital twins are getting a lot of attention in the context of machine learning. What will happen to it?

I believe the concept of digital twins is exciting and can significantly impact. Earlier, most analyses were done using simulations that did not align much with real-world settings. However, with digital twins, we can run many valuable simulations to study multiple processes in a system. Thus, by having better and constantly updated data, digital twins can have great potential to improve products and processes. Soon, we hope these models will be used in designing devices with limited memory to run multiple tests for creating robust models.

Some people think that artificial general intelligence is the true north star of AI. What do you think about this?

I am not very opinionated about artificial general intelligence (AGI). I have seen numerous AI techniques that have had an incredible impact in their respective fields. However, I have never heard of an ongoing AGI project that has led to significant research impacts. Nevertheless, I hope that the ongoing research project titled Gato at Deepmind can be highly impactful and may even initialize an entirely new field in AI.

What advice do you have for those who want to work in AI research? What are the most efficient methods of progress?

The best way to progress in an AI research career is to work under leading researchers in the area. Thus, the foremost task for those planning a research career in AI should be to identify the people working in the domain and to approach them for collaborations. Most of these researchers would be a part of a leading industry lab or an academic university, and their research projects may involve long-term milestones. Therefore, one should be ready to commit a few years to work under these researchers and understand the foundations of performing good research. Moreover, I would suggest not simply following the trends in research as these trends keep shifting, making it impossible to do quality research in this ever-changing environment. Lastly, one needs to have a deep mathematical knowledge of implementing ML models to pursue a research career. Thus, I would suggest they build a strong background by reading multiple books and viewing lectures from leading instructors.

Could you provide a list of notable academic books and journals in AI?

There are several books that one can use to have a strong foundation in AI, such as:

• "Pattern Recognition and Machine Learning" by Christopher Bishop.
• "Machine Learning: A Probabilistic Perspective" by Kevin P. Murphy.

For having a good foundation in deep learning, one can refer to "Dive into Deep Learning" by multiple authors, including Alex Smola.

For journals that include top-tier publications, one can refer to the journal of machine learning research (JMLR), journal of AI research (JAIR), transactions on pattern analysis and machine intelligence, etc. Moreover, we can access a majority of quality research by reading the proceedings of leading conferences.