Branch: B.E Computer Science + MSc Economics, BITS Goa
Batch: 2015
Current work: Pre-doctoral researcher at Google research in the Natural Language Understanding (NLU) group
Q. Can you tell us a little bit about yourself and your current interests?
Right now, I’m working on multilingual representation learning, with a focus on low resource languages. At a very high level, we work on language models which can do well on languages that do not have extensively available text, unlike models of English. I hold a keen interest in multimodal learning as well. Multimodal learning tries to bring all of the modalities into one space. I believe that's a more natural extension of how humans evolve. Others at Google research work on vision, language, and speech separately.
Q. Can you explain the concept of pre-doctoral research?
It's an opportunity for fresh graduates to get a research experience before they make decisions for further studies like an MS or a PhD. That way BITSians are lucky because of the thesis program. In fact, dualities are luckier because they get an extra year, as compared to their peers who get six months. Our college doesn't have many rigid constraints on taking thesis at different places so I could have a year of work being done at one place and in the field I wanted. The predoc program also enables one to get exposure to the research problems which have a significant impact in the years to come, helping one to craft their interests in those directions.
Q. When you were in college what decisions did you make for going into research in your field?
I did my PS1 at Liveweaver Pune. They were working on building NLP systems for a voice recognition software which gave me a brief idea of what NLP is. Being a dualite, my CS courses started in third year and that was when I started talking to a few seniors about placements, what to do next and how their job experiences were. I realized that the concept of a software engineering job wasn't very appealing to me. Moreover, I had a few seniors who were pursuing their masters/PhD, who were inclined towards research, and speaking to them got me thinking about the option of research. 3-2 was the time I applied for scholarships like DAAD and Mitacs but unfortunately, I didn’t get selected, probably because I did not have enough research experience at the time. In January I started applying to a few colleges in India and was eventually fortunate enough to land a summer internship at IIIT Hyderabad. My project was on generating code mixed data.
That was my first proper research experience and when I came back, I was further convinced that I wanted to do research. Although I found academic research a little slow-paced, I was still intrigued by it and I enjoyed the process of being handed over a problem statement, reading papers, and thinking of ideas to implement. I personally enjoy reading papers the most, as it gives me a lot of ideas.
Q. How did you build your resume?
When I was applying for a summer internship after 3-2, my resume was honestly not the best. I had done an informal project in economics and my PS1 project at the time. I still applied and fortunately did my summer internship at IIIT Hyderabad. When I came back in 4-1, I took up a software development for portable devices project under Prof. Sreejith B and continued to work on it for the whole of 4th year. The project was on emotional awareness for healthcare. We were trying to model the emotions of a patient based on visual cues, voice cues, and physiological signals. The goal was to help the patients communicate with their doctors and nurses on how they were feeling. We experimented with quite a few things and although we didn't get stellar results, it was much more of a learning experience for me because I implemented several classifier models and algorithms.
In 4-1 and 4-2, my resume was more substantial. My main projects were my summer project and my SDPD project which I was carrying out on the other side. I had taken ML and Natural Language and Fuzzy Logic (NNFL), so I included those as well. I did not have any formal SOPs. I had an economic SOP which I included because the work was more on data analysis. I was also the TA for applied econometrics where I was teaching python data analysis using R and Python.
Q.How has college helped you? Were you part of any clubs and departments?
I was an active member of the music club and it took most of my time. I think BITS Goa now has a society for artificial intelligence and deep learning which I think is a great initiative because they're doing a lot of cool stuff there. Unfortunately, it all started after I left college. I think the best way college has helped me is by connecting me to my brilliant seniors and batchmates. Moreover, I think the thesis option in BITS is one of the best things ever as it provides real research experience which is so valuable! However fancy the SOPs and DOPs look on your resume, doing a thesis is proof of concrete research work. Companies like MSR and Google AI have pre-doctoral fellowships and RAships for fresh graduates and the BITS thesis program gives us a major edge over others. So, a lot of BITSians have an edge over others applying for jobs because they already interacted with the researcher for six months or even a year.
Q: What is a corporate thesis and how does an aspirant get one for their thesis program?
Microsoft Research is well aware of the thesis 2 program that BITS Pilani has to offer since a lot of students approach them for the same. After going through the researchers' profiles I shortlisted three whose work was in line with what I wanted to take up. I emailed the three researchers/scientists and followed up at least three times. I eventually landed an internship with Sunayana Sitaram. Although it was too early for her to decide on interns in March, I persisted on taking an interview before since March was the deadline for college. The key here is persistence. The reason I went for a corporate thesis was that I wanted a research internship with some amount of funding. As far as interviews were concerned, mine was easier compared to other research fellowships. I do not think companies expect us to be experts right off the bat and they can figure out when a candidate is lying in his/her resume. So make sure that the projects you list are actually the ones you have done and can explain when asked.
The first six months of my Microsoft Research internship have been the best learning experience of my life and also the most rewarding. They're very flexible and when you have all your time devoted to doing research, you can do a lot! In the first month, I was working on the previous intern’s project; he had been working on spoken keyword detection, so I made a web app for it, for them to present it in a Microsoft hackathon. After that, my own project started. I think for the first one to one and a half months, I had a problem statement in mind, centered around transfer learning techniques for low resource NLP. The first one and a half months were spent just reading a lot of papers and learning about the recent architectures making a wave in the NLP community like Transformer, BERT etc. The reading phase ended with a presentation I gave to the entire NLP group because a lot of us were still unaware of the details of the new architectures. Also, we had a conference coming up in December. It's called ACL. It's an A-star conference in NLP. The presentation was what gave us the idea of what to work on for ACL. We wanted to work on building a benchmark for code mixed NLP similar to the GLUE benchmark for English. I worked on that for the next three to four months. I also worked on building a new data set for it because there was no existing data set for one of the tasks. I could actually get two papers out of this work at good conferences like ACL and LREC Workshop. After that, I continued my thesis and that was that. So for the first six months, I was working on code-mixing, but in the next six months I shifted to multilingual learning
Q: What technical and soft skills do you feel are most important for juniors to pick up in college?
It’s different for a research profile versus a developer profile. For placements, it's DSA. Being very proficient at data structures and algorithms, really, really helps you sail through placements. By proficient, I mean, really really proficient. So even in Google SWE roles, I think they just have some rounds mainly focused on DSA. If your aim is for a job in IT, then you have to be very proficient in it, especially because a lot of people start with competitive coding in their first year or even before that nowadays. So the competition is real.
For research, I would say, you've got to take up projects. The focus is less on your existing knowledge, and more on whether you have conducted successful research. Even when you're applying for the pre-doctoral researcher position, you don't necessarily need to have experience in the exact sub-field you want to work in. So it is okay to be interested in one subtopic and have a paper in another if they're not completely orthogonal. A paper shows that you have the potential to conduct research and you have gone through the entire process. So the best way I think is to take up research projects with good professors in your university. Concentrate on implementation of the project, i.e get your hands dirty. Proficiency with ML frameworks like PyTorch and Tensorflow really helps on the way forward.
Disclaimer: The points given above are the views and steps taken by the individual. They are not fixed steps and guidelines to base your college upon. Our hope is to inspire students so they can take the necessary steps hereafter. We hope you like it!
Comments