It’s been a while since I’ve posted, and a lot of things have happened since May, 2022. It would seem that time at IITB passes by very quickly. I’ll adopt a month-by-month posting format here instead of the usual list-of-courses format.
I came to the campus for the first time since the start of my program. In the process I met some of my batchmates and my advisor as well. At this point, I was the sole student on the influence maximization project from the previous semester that would eventually get accepted to WSDM. I continued to push along slowly, trying to get results on various baselines and comparing them to our method(s). We had missed the previous conference deadline in May (due to the passing of my grandfather, among other things) and I wanted to make sure we wouldn’t miss the next one. After 3 weeks, I decided to head back home and continue working from there.
I came back to campus on July 19th, and switched hostels. I had gotten some solid results in the past month and my advisor (and our collaborator) decided to submit to a conference. I also got my first taste of illness and the on-campus hospital gauntlet, having been put out of commission for 4-5 days. I registered for combinatorics (CS604) this semester, with the goal of improving my proof construction, having struggled wth proofs of submodularity (and gamma-submodularity) in the research project back in April. While I did gain some intuition on proofs and proving methods in different scenarios, I wasn’t really that good at it.
As an aside, a short description of CS604: the course, run by Prof. Sundar, was divided into three segments - graphs (use Douglas West’s book), hypergraphs/set systems and discrete probability. Stasys Jukna’s Extremal Combinatorics was the main book for the course. Each segment had an assignment associated with it, and we had try to prove the results by ourselves, and use lecture videos after some attempts. There was no real “instruction” in this course - it was meant for solving problems and then discussing solution approaches in class. Three quizzes were conducted and two sets of homework problems (not really related to the topics at hand) were given out. There was no mid-semester exam, and the end-semester exam held an ominous 50% overall weightage. Take this course only if you’re theoretically inclined and/or have the time to make up for it by trying out many proof based questions from books/other material.
Things had gotten heated up in the lead up to the conference deadline. We were initially chasing AAAI, but then decided on WSDM for a more topic-targeted submission. My birthday came and went quickly in the midst of work (grateful to my friends for celebrating it regardless). I also had a rather unpleasant TA experience close to the deadline but managed to weather it. We finally submitted the paper very close to the time deadline and I took some well-deserved time off after this. Later this month, I housed my parents over at the IITB guest house when they came to the campus for the first time, which felt amazing.
I was put on a project with an applied scientist on the research division of a well-known e-commerce company. The original idea involved improving recommendation systems with the help of a subset selection algorithm. However, I came to realize that the codebase left to us was not in a very workable state, and was hard to understand. We eventually pivoted to a method based on a mix of curriculum learning and matching networks, both very interesting ideas in their own right. We planned to submit to WWW the next month, but as fate would have it, we could not continue due to administrative issues. For what it’s worth, it was a good experience and the AS was very helpful and guided us well.
The most notable incident from October (or what should have been, anyway) was the notification of acceptance to WSDM. What should have been a month revolving around celebrating the acceptance and putting my head down in other projects, instead became a month of neurochemical mind trickery and intense emotion. I’ll spare the details on the weird experience, which lasted into the end of the year. The semester concluded in November with the end-semester exam for CS604. Some of my batchmates and I also applied for Google Research Week, which was due to happen in January.
My guide and I had discussed more ideas for my thesis, including an adversarial flavour to influence maximization, but I did not progress much at the time. Eventually the time came for my BPS (in November), or semester-wise assessment of progress towards my thesis. My committee consisted of two professors from the CSE department, both working in varying degrees of theory. I presented my WSDM work to them and explained influence maximization in brief. In return, the professor who worked in game theory suggested ideas for opinion-dynamics based influence spread in social networks. With this, I went back home at the end of the month.
Other than attending weddings and wedding-related events this month, I also got the chance to attend the IndoML symposium at IIT Gandhinagar. It took place over the course of three days where a number of speakers from various institutions delivered talks. A poster presentation was also conducted on the first day, where (mostly) PhD students (some from outside India) presented their work done so far. I met with my guide here as well, who gave a tutorial on submodularity alongwith a collaborator professor from UT Dallas. I returned to campus after this, and got some work done. Mood Indigo was a refreshing change at the end of the month - I attended a Sunidhi Chauhan concert for the first time too. Some of us also received an accept notification for Research Week around this time.
I was parachuted (drop-shipped?) by my guide into a project on learning resource-efficient mixture models for clustering on the first day itself. The goal was to carry this project to ICML, whose deadline was at the end of the month. I was given the responsibility of the experiment pipeline and was to obtain results on a completely different set of experiments than what was originally thought out for this project. This we did over the course of the month, and I had help from one of the first-authors of the paper, to whom I am grateful.
I registered for CS728: Organization of Web Information, which happens to be a NLP research course organized (heh) and run by Prof. Soumen. Prior experience in NLP is not mandatory but recommended, and that in machine learning/deep learning is preferred. In the initial classes classical sequence models were discussed, progressing to word embeddings and representation, basic neural models for sequences, and sequence-to-sequence models and attention. We also covered many transformer-based models. In later classes we covered applications such as NER, knowledge graphs, knowledge-graph guided question answering, KG representation, multi-task language models and large language models in general. Oh, and one class was devoted to graph neural networks (which I am glad for, because that one fun question on GNNs in the end-semester exam really set me up). I was lucky to have two batchmate friends who were already pretty good at NLP with me in this course.
We submitted the ICML project close to the deadline (again), and then I went with one of my friends to attend Research Week in Bengaluru. We were put up by Google at a hotel near the venue. There were talks by the heavy-hitters of their respective fields, and I had insightful conversations with researchers such as Partha Talukdar and Prof. Balaraman Ravindran during socials. We were quite grateful to Google for their hospitality and expert management of the event(s).
By this time, my e-visa application to Singapore was ready and I submitted it to the travel agency. I worked with multiple professors and the CSE department to obtain leave permissions around the time of the conference. With a small detour to Kolkata for a family function, I headed to Singapore that Sunday to attend WSDM.
My experiences at Singapore warrant a post on their own, given that it was my first time outside national borders. It was fun in general and I made a few friends at the conference. My paper was due to be presented as a poster on the second day, and I had to explain (and defend) our design decisions in front of other people for two hours. It was a surreal experience. If you’d told me three years ago that this is what I’d be doing in the future, I would not have believed it. I also attended some workshops and tutorials on the following days and ice-breaker events run by the organizers. While my guide did not attend, the collaborator professor from IIT Delhi did, and I met with him and his student. His other paper even made it into the WSDM top 10, which was pretty cool.
I was back on campus by the first week of March, with another minor detour to Bengaluru to complete the family function circuit.
In short, these months passed mainly in course related stuff (quizzes, end-semester exam), discussion of other projects, presenting my WSDM poster at a department research symposium and participating in the ICML rebuttal. In April, we received news of the accept, which was a pretty big deal.
Now I’m back on that multi-project grind. More for a later post!