Dr. Maunendra Sankar Desarkar is a part of the Natural Language and Information Processing (NLIP) Research Group at the Department of CSE, IIT Hyderabad. Before Joining IIT Hyderabad, he worked for Samsung Research India Bangalore and Sybase Inc. (a SAP Company).

His main research areas are Natural Language Processing, Information Retrieval, and Machine Learning. 

INDIAai interviewed Maunendra Sankar to get his perspective on AI.

What inspired a software engineering professional to pursue a Ph.D.?

Doing a PhD and joining academia has always been a goal. It started from my BE/BTech days, where I developed a keen interest in how the different technological components work - creating from the fundamental aspects. Joining the industry exposed me to how those technologies can be used and how basic concepts which power the technologies remain the same, whereas the technology may change. It made me even more interested in exploring the concepts and developing newer algorithms, frameworks, problem statements, and applications that solve real-world needs. PhD was the next logical step to achieve this.

Tell us about the initial challenges you faced during your research. How did you overcome them?

I joined PhD after a two-year stint in the industry. Two years is a long time, as it happens in research. Many things had changed; newer techniques, results, frameworks, and applications had come up. I had to spend a considerable amount of time going through research literature to update myself. Then, when I selected a problem statement and developed an algorithm with some sound theoretical foundations, the empirical results needed to be better. I had to look at the data and the intermediate and final results produced by my techniques and figure out how we could do better so that the method is theoretically and empirically acceptable. 

Finally, we could showcase our algorithm's promise by carefully analyzing the algorithm and the results. It got accepted at a very reputable venue with extremely encouraging reviews. It shaped the path of my PhD research too. It took more than two years to get my first publication. I have published many papers after this. However, this publication will remain memorable as it taught me to deal with failure, persevere, and work hard.

Tell us about your research area and your research contributions.

I work mainly in three areas: Recommendation Systems (RecSys), Information Retrieval (IR), and Natural Language Processing (NLP). In the last three years, the focus has been more on Natural Language Processing as there is a lot of interest in this area. My other interest areas, Recommendation Systems and Information Retrieval, also benefit from the advancements in NLP. In the space of NLP, I am focusing on responsible NLP - how to enable language technologies for resource-scare languages (most Indian Languages fall in this category) by doing superior cross-lingual transfer, making the digital space cleaner by detecting hate or offensive content, and also ensuring any machine-generated text is free from toxic or harmful or derogatory content.  

Another aspect is to look at explainability in all the tasks we handle so that one can reason the output produced by the system. A primary focus is to see how NLP and other AI techniques can be used for the social setup - this includes both developing solutions relevant to the social sector and also ensuring that the solutions or understandings can be accessible, usable, and deployable by any stakeholder - a major MNC, a startup, or even an NGO. For example, our works on multilingual NLP aim to make natural language content accessible to a larger audience that speaks different regional languages. With an NGO, we are developing an AI-enabled job portal for disabled youth. We take special care to reduce the memory footprints of the developed models so that even small players in the business or social space can deploy these.

You would have reviewed numerous research articles; in what areas are scholars and students deficient?

I have been in the reviewer panel for many good journals and conferences. I am happy to have reviewed a good number of excellent articles. However, many articles need considerable rework. There are times when the technical contributions could be more present. On the other hand, where technical contributions might be there, there are other areas for improvement. The deficiencies that I observed in this category are of multiple types. The authors often list what they have done without motivating the problem and backing the solution strategy through intuitive justifications. In many cases, the algorithms proposed are for a well-known problem. 

Here, the authors are expected to explain why there is a need to have a new approach - what the drawbacks of the existing techniques are, and where the proposed solution fits in that landscape of existing related work. Often, the experimental results need solid baselines, or the insights from the experimental results need to be analyzed. More than simply saying that our method has better numbers is required. In many cases, the presented experimental results are minor improvements over the compared methods, where having correct predictions for two or three additional/more instances would be sufficient to have those improvements. Researchers need to understand these aspects while preparing and submitting papers for publication. Going through good research papers from prominent venues helps address some of the above points connected with presentation and writing style.

Who should consider pursuing a career in data science? What does the future hold for someone who embarks on it now?

The coverage of Data Science is now broadening. Many applications are coming under this broad umbrella. The inputs and the targets of these applications can be from different modalities - text, audio, image, video, or a combination of more than one. For any task, the starting step is the processing of the input. Hence, one needs to be familiar with how to deal with these different types of data. Next, any technique can be applied on any data, and it will produce some output. One must be curious and willing to understand why this specific output is produced. It will help in understanding whether the underlying method needs to be changed, even if it has reasonable performance and a further improvement is possible, or if the methods should not be tried at all as there can be a broad range of inputs where it will not work. 

If these justifications can be obtained before trying out the methods, they can make a data scientist a real asset. This role is called Data “Scientist”. So, one needs to be curious to know what is happening inside the methods, and how the intelligent predictions are made. The job of a data scientist should not end at merely putting different available components together, and simply reporting the metric scores. The role requires good programming skills, mathematical reasoning, knowledge of statistics, and a curious mind. The technology landscape is changing, and technology professionals need to be adaptable. Hence, relying on knowledge of tools will not be sufficient. One would need to understand the basics in a top-down and bottom-up manner to perform well in this changing and challenging environment.

Any candidate aspiring to take the data scientist role should possess the above skills. We will now see more and more codes and test cases generated by copilots. However, when the generated codes do not perform well, or when it is essential to understand whether there could be a further boost in the performance of the systems, for understanding and quantifying the biases (if any) produced in the solutions, these insightful people will have a lot of value in the industry.

What advice would you offer to someone looking to get into AI research? What is the best way to advance?

Amy research requires a curious mindset and perseverance. Apart from these two, AI research requires a specified set of skills such as concepts of mathematics - linear algebra, probability and statistics, and optimization. As data is an integral part of any AI system, establishing the superiority of research contributions needs to be validated on large datasets. As a result, there is a lot of emphasis on implementation and working with large-scale data. Consequently, one needs to be a good problem solver and possess good programming skills. Engineering skills as well as algorithmic skills are both necessary for being a successful AI researcher. In addition, candidates with good written and oral communication skills will have an advantage as it will become easier to interact with peers and collaborators to discuss solution ideas in this rapidly changing field.

What are some essential research articles and books that inspired you?

My inspiration for research was through continuous exposure to the basic concepts of different domains. However, several research articles had me in awe regarding how the problems were explained and the solutions were motivated and described. The book “Introduction to Information Retrieval” by Manning and Raghavan and Speech and Language Processing by Jurafsky and Martin exemplifies how complex topics can be presented lucidly without disregarding technical rigour.