Ankit is a PhD Student at Computer Science and Engineering department of IIT Delhi. He is advised by Dr. Parag Singla and Dr. Mausam. He completed his M.Tech from Department of Computational and Data Sciences, IISc Bangalore. His primary research interests are Artificial Intelligence Planning, Reinforcement Learning and Probabilistic Graphical Models.
- His current work
- Differences in research environments between Indian and foreign universities.
- What young graduates wishing to pursue ML should focus on
ML India: We’d like to start by understanding your background and how you got into the field of machine learning. Was there any particular instance that sparked your interest in the field?
Ankit: I was very interested in maths during high school. Even though I didn’t have any experience of coding or computer sciences during high school, I gradually started to see a relation between math and computer science after taking up computer science engineering at Punjab Engineering College in 2006 for my undergraduate studies. I did face some problems initially while trying to learn the basics, but I eventually learned languages like C and C++, and built some interesting applications in the first couple of years. I then shifted my focus towards extracurricular engagements. I was a part of an organization called Les Amis, where we used to conduct and organize city level events. It was one of the most vibrant and well-formed organizations that I’ve worked in. Since I was doing well in both academics and extracurriculars, I was, at one point, inclined to pursue an MBA. I cleared some entrance exams as well but was not able to crack the top IIMs. I had simultaneously also appeared for GATE and was able to rank in at 123. That year, I had applied only to IISc Bangalore and also got a call from SERC, but after much deliberation, I decided to join Samsung India Research Labs instead. After getting some experience at Samsung, I felt the need to advance my knowledge.
After resigning in February 2011, I decided to pursue a Master’s degree at IISc Bangalore. I got into the Department of Computational Sciences (CDS) which, at that time, was a part of Supercomputers Education and Research Center (SERC). Most of the courses they had were around mathematics and computation, which aligned with my interests. Also, I immediately noticed the difference in the academic environment at IISc from my experiences at other engineering colleges. IISc fosters a strong research environment, and my interactions with people there instilled a lot of enthusiasm in me and got me really interested in spending time there and pursue a career in academic research. They also had courses like ‘Game Theory’ by Narahari, which are generally not floated at a lot of places in India, which was another reason why I decided to enroll myself at IISc. Before I joined, I had a couple months to myself, so I took a break from academia and interned at the Nehru Learning Centre for Children and Youth(NLCCY), Ministry of Culture and Youth Affairs, where I worked towards increasing accessibility to cultural heritage institutions like museums.
At IISc, I went through a number of mathematical courses during the first year. We had a good degree of flexibility on what we wanted to learn and the classes were very enriching and interactive. The coursework also helped me in brushing up my knowledge on basics of statistics and linear algebra. As a part of the computational science coursework, I was studying some courses related to data analysis and visualization but I had never done computer systems oriented work in my career. I thought that it was a good time to experiment and explore new fields of study and I started working in the Cloud Systems lab under Prof. S.K. Nandy and Dr. J. Lakshmi.
This lab was set up just 6 months prior to my joining, so we built a lot of cloud infrastructure for it and worked on many optimization algorithms. I also managed to pull a national and an international publication which boosted my confidence in the research domain. It was at this point that I realized I wanted to learn more and work in the theoretical and algorithmic areas, and that I should pursue a Ph.D. I chose to enroll at IIT-D for my Ph.D. where I worked with Dr. Naveen during my first year, and picked up some theory courses. I also picked up a machine learning course with Dr. Parag Singla since I believed algorithms and ML go hand in hand. After doing a couple of projects with Parag, I realized that I can actually proceed to work with Parag Singla in the field of AI. Mausam also took part in some of our discussions and meetings, and subsequently became my co-supervisor.
ML India: What is your current work focussed on?
Ankit: Currently, I am working on AI/ML projects with Dr. Mausam and Dr. Parag. Broadly, I am working on exploiting symmetries in artificial intelligence and machine learning problems. We believe that there is a lot of symmetry in AI and ML problems that can be utilized to reduce the size of the problems and make the algorithms more efficient.
I have applied this to two areas. The first is Planning and Reinforcement learning, which Mausam worked on during his PhD. We exploit the symmetries of states and actions in Markov Decision Processes(MDPs) to merge states and actions together which can be used by MDP solvers. We are working on the state-of-the-art algorithms, specially the Monte Carlo tree search, which was a key algorithm in AlphaGo. We worked on UCT from this category, abstracted state space and actions and built ASAP-UCT algorithm, and improved its efficiency. We also got a couple of research articles published on this work.
The other thread I am working on is exploiting the symmetries in probabilistic inference. This is Parag’s core working area, also called as Lifted Inference. We work on exploiting symmetries to reduce the model size, which is termed as ‘lifting the domain’, and hence the name. When a domain is lifted, various traits for same variables get merged and the algorithms become more efficient.
I also collaborated with two bright undergraduate students, Aditya Grover and Ritesh Noothigattu. Aditya and I worked together on a couple of projects which resulted in publications in top-tier conferences like AAMAS and IJCAI. Our work in this area got really good reviews from the community and we got a ‘Best Paper’ award at the STAR AI (Statistical Relational AI ) workshop where we worked on exploiting Contextual symmetries in Probabilistic Graphical Models.
ML India: How do you plan to conclude your Ph.D.? Will you be working on a new project or will you extend your current work?
Ankit: I have approached most of my work from a theoretical point of view up till now. Although it did have some practical aspects, it was majorly focused on improving theoretical features. I believe there are 2 ways to do a Ph.D. The first is to pick an application, try and solve problems related to it, and then making it more efficient using theoretical tools. The other way is to take up theoretical tools, work on improving these tools and then check for its applications. I followed the second path. For the past four years, I have been working on improving various algorithms and techniques to increase their applicability and efficiency. My work in the next 1-1.5 years will mostly be directed towards devising useful applications of my theoretical work. I will try focusing on exploiting symmetries in Computer vision and NLP applications.
ML India: Have you thought about what you will be doing after your Ph.D.? Do you have an inclination towards either academia or the industry?
Ankit: I am open to both academic and industrial opportunities. Research labs like MSR and Xerox Research which are doing amazing work in the industrial domain. It depends on the kind of work I get to do. I did an internship in an academic institution last year and got a good idea of the work environment. To get a better idea of how things work on both sides, I am planning to intern in the industry this summer.
ML India: What was your work focused on during your 3-month internship at Oregon State University?
Ankit: I worked with Prof. Alan Fern and was part of the AI Planning and RL group. At OSU, I extended my work on exploiting by symmetries to abstract state space features. I worked on exploiting general abstractions which included temporal abstractions (time) and space-based abstractions in tree-based framework. I started to work on developing a MCTS based MDP solver and am still working on that.
ML India: Did you observe any differences in lab environments and operations labs at OSU and at IITD?
Ankit: I didn’t find a lot of difference in terms of the advising style of the professors. But one of the key differences is that the number of Ph.D. students at OSU is much greater. Our group at IIT-D is pretty new and undeveloped. I am one of the first students to work with Mausam and Parag and all of us are constantly trying to bring in and execute new methods to attract more students. When I got to OSU, there were already 8-10 students working in the planning and RL area. So I got to interact with new students as well as with seniors and thus, the entire environment was more palpable. They also had reading groups where students working in common areas could get together to read and discuss papers from top conferences. The major reason why such reading groups have failed in India is because of less number of Ph.D. students. We only have a total of 5-6 Ph.D. students who are all working in different areas of AI and ML applications. But after I returned from the US, I decided to hold such reading sessions to discuss papers which have a little overlap of some or all of our working areas. This way we’ve been getting to learn about other fields and also about the work going on in our own research areas. It’s been going great so far.
Ankit: Nowadays, there is an easy accessibility to ML courses on the web. Students can definitely begin to learn about ML through such courses. Also, they could play around with openly available tools, and interact with people and professors on web platforms to know more about the field of ML. This would be a really good way for an outsider to get acquainted with ML.
“One cannot undermine the importance of mathematics in ML” -Ankit
But for an undergraduate who wants to actually pursue ML, I must say that a strong mathematical foundation is necessary. For any undergraduate student, or someone who has already graduated, I would suggest brushing up the math fundamentals. Some courses I would suggest are- linear algebra, probability, statistics and a little calculus. Another course on Advanced Numerical Optimization by Shirish Shevde, is a great course to learn from. These courses will help them to understand how the tools are working through a mathematical lens. They would eventually see that the beauty of machine learning lies in algorithms. I would restate that math is an integral part of learning and working in the field of ML. One cannot escape it, so better learn to embrace it.
ML India: Does your research group do any industry collaborations as well?
Ankit: Yes, Mausam has had a number of collaborations with Google, Bloomberg etc. I believe he is working with an educational startup as well. Manik Verma, a researcher from MSR, is also working as an adjunct faculty at IITD and has a couple of PhD students working with him. Their work is focused on recommender systems which is very well oriented to the industrial applications. I do feel that we need to have more such productive collaborations.
ML India: Machine Learning and datascience have now become a buzzword globally. Do you think this hype of ML has brought about any positive change on the Indian industrial work front?
Ankit: I think it has already brought around a lot of change in our everyday lives. Today we are generating data at an enormous rate through various electronics devices. It has become really important to make sense out of this data and harness it properly. People in the industry are actually applying various machine learning techniques on this data to come up with useful results. ML is constantly changing how we are approaching our life. The industry is leveraging the new technology, and constantly trying to build better products for the customers to make their lives easier and more efficient.
India, however, has still not gotten into utilizing ML completely. There are new startups which have dipped into the app development domain, but not a lot of work is happening towards improving the technology. We need to come up with the next awe-inspiring innovation and I think that an effective collaboration of academia and industry can help achieve that. The academia is doing well here and can help the industry by letting them know about the best tools that can be used to develop new technology. Companies like Google Deepmind are publishing papers in top tier conferences and are the biggest publishers from the industry. This year alone they published around 14-16 papers in NIPS. I want something similar to happen in India as well, where a company with a huge potential can get a backing from one of the big guns, and produce fantastic new research. I believe the next 10 years belong to AI and India will have a big contribution to make in its development.
ML India: Apart from academics, are you involved in any extracurricular activities ?
Ankit: I have been a cycling enthusiast since my Master’s. I started a cycling group- BTwin Diaries at IITD, and we cycle to historical sites in Delhi every weekend. It’s a good break from the regular academic tasks. I also am a cricket enthusiast and like to follow Harsha Bhogle and Rahul Dravid. I feel that Ph.D. is a long journey which should not be completely oriented towards academia. I believe that students should not let go of their hobbies while pursuing studies, as they provide a good balance and actually help you perform better in academics.
ML India: Thank you for your time Ankit. All the best to you from ML India.