Starting from 10th September, I am working as a research intern at MIDAS lab at IIIT-Delhi. MIDAS is actually an acronym that stands for Multimodal Digital Media Analysis. Headed by Dr. Rajiv Ratn Shah, the group currently conducts research across a broad spectrum of domains such as speech recognition, lip reading, musing generation, social media analysis etc. (these are the ones I know 😁). MIDAS is a young research group, it started out in early 2018. Notwithstanding the short time since its inception, MIDAS has distinguished itself in top-tier conferences like EMNLP and ACM Multimedia.
I first came to know about MIDAS from a friend I met at PyData-India 2018. A few weeks later I came across a congratulatory post for EMNLP acceptance of a paper from MIDAS. This was the dealmaker for me. So here was an AI research group very far from my campus that is churning out papers at prestigious conferences. With this thought, I immediately sent my resume to Dr. Shah for a research internship position at MIDAS.
I got an interview call soon after I sent my resume (I guess after 3 days). The interview was conducted through Google Hangouts by Dr. Shah and my would-be mentor Yaman, who incidentally is an alumnus from NSIT (a slightly unfair advantage 😁). The interview was mostly centered about my personal projects, past experience in Deep Learning and especially my internship project at Samsung Research. The interview was a breeze and was more like a frank technical discussion. I was then told about some recent projects that the lab is currently working on and the progress they have made in related areas.
Funnily Rajiv Sir asked me about what MIDAS stands for, and I embarrassed myself by fumbling about the myth around a Greek king named Midas whose touch could turn things into gold. I suppose this is Rajiv Sir’s go-to question since he asked the same question to a couple of co-interns. So it would be helpful if you memorize this answer 😁.
Currently, serious research in Deep Learning requires a massive amount of computation. Any non-hobbyist work in this field cannot be conducted without GPUs. MIDAS currently has access to a powerful machine at NUS Singapore (called Weisshorn) and another one at IIIT (called Falcon). Together these machines satisfy the requirements of the whole team. But as MIDAS expands, it will have to obtain more machines.
This is the part that I love the most about MIDAS. Prof. Rajiv organizes a monthly meetup (I got an opportunity to attend one of those on 13th October) where everyone talks about the task they are working on. One has to explain the approach he/she finds promising and can get feedback. I feel this feedback loop is extremely important since in research it is easy to lose track of our work while developing a solution that we think is promising or miss out some intricate detail that may be critical to your idea. And this is where working in a group is advantageous. It ensures that:
- One doesn’t get stuck. Thanks to the feedback cycle.
- One is up to date with state of the art ideas. It’s quite difficult as an individual to keep track of new papers and breakthroughs, especially in a rapidly developing field like machine learning.
Even if I ignore these advantages, these MIDAS meetings ensure that everyone has a sense of belongingness towards the group and makes me feel more like a collaborator than a student intern. This is a model that every Indian research group should try to emulate.
Finally I would encourage students who are interested in machine learning to apply as an intern/research assistant at MIDAS. If you are a student at NSIT, then you are in luck. You would certainly feel at home since a significant proportion of members are from NSIT.