Interview with Chetan Arora, Associate Professor, IIT Delhi, India

Chetan Arora is an Associate Professor with Computer Science Department at IIT Delhi. He received Ph.D. in Computer Science and B.Tech in Electrical Engineering, both from IIT Delhi in 2012 and 1999 respectively. He did his Post Doctoral Research with Prof. Shmuel Peleg from 2012-2014 in Hebrew University, Israel.  Prior to joining IIT Delhi, he was an Assistant Professor with IIIT Delhi. Chetan has published more than 35 papers in top computer vision journals and peer reviewed conferences. Prior to returning to academics, Chetan has spent over 10 years in industry, where he co-founded 3 startups, all working on computer vision products coming out of latest research ideas.

Chetan has served as an area chair at ICVGIP 2016, program chair at NCVPRIPG 2017, Workshop on Assistive Vision held with ACCV 2016, and Workshop on Computer Vision Applications held with ICVGIP 2018. He is also the program chair for the upcoming ICVGIP 2020. Apart from his academic research, he is also committed to applications of computer vision for persons with disabilities.

Q1. Please share your current research work vis-a -vis relevance from Machine Learning context.

Reply: My research program deals with how to make current AI, and machine learning techniques more trustworthy. And when I say trustworthy, high accuracy is only one aspect, and many times, not the most important aspect of the trustworthiness. To be trustworthy, the systems, must be robust to noise, and environment distortions in the input. They must be rational, and logical in their prediction in terms of agreement with our prior knowledge. Reliability, defining a particular specification of a machine learning system, and then functioning predictably within those specification, is another important aspect. Interpretability, and explainability as well as Fairness, Accountability, and Transparency (FAT), are other important aspects which researchers in this area are interested in. In my research program, we have looked at robustness, and incorporating prior knowledge in the traditional machine learning techniques, using MRF-MAP framework. With the popularity of deep neural networks for solving most computer vision problems, we have started to look at adversarial attacks, and reliability aspects of such systems as well.

Q.2. Would you please share any of your impactful work with us?

Reply: Our research group has been at the forefront in developing algorithms for posterior inference in MRF-MAP problems. As indicated above, the formulation is an important aspect for imparting robustness, and including prior knowledge in the predictions of a machine learning system. There are various sub-problems within this largish domain, such as when the number of classes are two or more, when the prior knowledge can be formulated as submodular functions or not. In almost all these sub-domains, the state of the art algorithm for the inference has been contributed by my research group. We have repeatedly shown over last few years that using these algorithms one can improve the predictions of state of the art techniques based on deep neural networks.

I also believe that apart from focusing on research and academics, our responsibility as the members of such elite institutions, is to also contribute to nation building and socially relevant problems. I have my personal interests in two such problems. In the first we are focusing on the breast cancer detection from mammograms. Breast cancer is emerging as the leading cancer among the women in India, and because of data centric nature of modern ML techniques, these computer-aided diagnostic tools often do not work successfully for Indian population. Through our collaboration with AIIMS, Delhi, we have been able to access a large dataset of mammograms. Our mission is to invert the situation, and develop next generation of AI tools, with state of the art performance on the Indian population, and let western world follow us instead. Our efforts have started to bear fruits, and some of our latest techniques are now state of the art for not only Caucasian but Indian population as well.

The second problem of social relevance that we are interested in is mobility assistance for visually impaired. India has one of the largest share of the visually impaired people, and the unstructured conditions in our country (such as dogs/cows on the streets, broken/non-standard pavements etc.) create a big challenge in their independent mobility. This hampers their professional and economic growth, and productive assimilation in the society. We are working towards a camera based wearable device which can assist such people to move safely and independently in the unstructured Indian outdoor scenarios. We have completed a first prototype of the device and are doing user trials for this.

Q.3. In your opinion, what are some of the most exciting areas of research in ML for students and upcoming researcher?

Reply: Without offending my fellow researchers, I feel that whole area of machine learning is much hyped at this time. So far, we have demonstrated good results on the applications, in which accuracy is the sole metric. For example, consider the face recognition problem, where one would like to search the face of a particular person, say a celebrity. As long as such a system has good precision, and recall it doesn't matter what faces it picks or misses. But as we go forward, we are looking at applications of machine learning where there are life and death decisions associated with the predictions. Hence, every error that such a system makes, or even every prediction that is made, must be carefully calibrated, and well understood. Hence, all the areas or sub-domains which gives us a handle on such problems are going to be very important.

Most of the deep learning techniques today are supervised and extremely data hungry. As we look to replicate the success of these in newer application areas, we are not going to have the luxury of large datasets. Hence semi-supervised, weakly-supervised, and un-supervised techniques will become more important. Many of these techniques will not be able to work without having a good data and knowledge representation, which will also be an exciting research area in near future.

Q.4. What challenges do you think our current student population faces as far as preparedness in ML is concerned?

Reply: The availability of variety of libraries and tools, as well as easy access to computation hardware has become both boon as well as bane for the machine learning research. I increasingly see a generation of our students happy at seeing themselves as expert in running scripts and satisfied at gaining superficial knowledge of the area or a problem. This makes me concerned that we are training a next generation of ML coolies, and not the innovators of future. There is no dearth of resources or information, what we only need is a change in mindset, and preparedness to deep dive. We need many more students who opt for PhD in these areas, working on fundamental problems, and make India a leading place to do ML research.

Q5. There are many online courses available in ML. Can you please share your suggestions for students so that they are benefited most by doing the courses in this area via online learning?

Reply: I guess the information/awareness of students would be much more than me on this topic, and hence will refrain from making any recommendation. However, I would like to share one of my observations. I have felt that MOOCs from many of the popular websites are often substantially toned down versions of the ones taught in universities. Hence, I would encourage students to not consider it complete, and stop after doing such MOOCs but also have a look at the course material which is usually available at university instructors' websites. There are often no video lectures but only slides available. However, after one has a basic understanding through any of the MOOCs, such lecture slides are also usually enough.