It’s been raining international interns at IIITH lately. This time, we tracked down Marius Dufraisse for a tête-à-tête. Did you know he has quietly been working at the Machine Learning Lab for over 5 months? Well, now you do!
‘An Indian experience’ wasn’t entirely novel for 22-year-old Marius Dufraisse, having toured Kerala and Tamil Nadu with his intrepid family a few years ago. “This meant that I wasn’t too afraid of coming here. I knew the food. And I knew what to expect about the weather, otherwise I’m not sure I’d have been here,” he says. As a Masters student of Computer Science at École Normale Supérieure (ENS), Paris-Saclay, Marius’ coursework required him to complete a 5-month internship in his first year. Advised by one of his teachers and a former CVIT researcher Kartheek Alahari to apply to IIITH, Marius says it was one of the best things he’s done.
ML Lab
Prof. Girish Varma who works jointly with the Machine Learning Lab and the Centre for Security, Theory and Algorithms Research (CSTAR) says that just like other faculty members, he too received many applications for internship this Summer. What stood out in Marius’ application was a reference made to a well-known paper that Prof. Varma had worked on previously. The research involved making large neural networks simpler and smaller. Marius evinced interest in this work, and requested to collaborate on it or something similar. The other plus point that worked in his favour was the fact that Marius was a fully-supported student at ENS, Paris-Saclay. “Marius didn’t expect any funding from us, and he was already a Masters student,” says Prof. Varma of the selection.
The Project
At the Machine Learning Lab, under Prof. Varma, Marius worked on slightly different problems though. He was assigned to work on “generative models”, or very simplistically put, generating realistic images with Machine Learning algorithms. “A human after being exposed to and having seen the world enough can start sketching the same, but the problem we explored here was if an ML algorithm can do the same, explains Prof. Varma. The team worked on facial images where given a dataset of images, the model could not only generate new images but left the researchers a greater degree of control over what was being generated. To understand this better, think of the AI photo editors that recently had even the celebs running in their pics through the old-age filter to see what they’d look like when they’re old. “A lot of solutions already exist but we are working on a newer set of methods where we can also understand how the model works. The current models use Generative Adversarial Networks (GANs). We’re doing it by using Normalizing Flows which give us more control over the neural network,” says Prof. Varma. In Normalizing Flows the neural network used is an invertible function. That is, it can not only be used to generate an image from an intermediate representation, but the reverse is also possible, allowing one to do things like modifying the image. For example, control the age, remove glasses etc. “Our research was mainly focused on designing neural networks layers that are invertible and more expressive, resulting in better quality images,” explains Prof. Varma.
Background
With both parents as high school teachers, and the father teaching Math, perhaps it doesn’t come as a surprise that Marius found himself “doing a lot of Math at one point”. “I read a book about Computer Science and it seemed really interesting. So I felt like I should try to push through studying in this field,” he says. Towards this end, after high school he enrolled himself in a 2-year program in Paris known as the Classe Preparatoire where the focus is only on Math and Physics to prepare for competitive exams. Getting into ENS, an institute acclaimed for the Sciences and Technology, he completed his Bachelors and is now currently pursuing his Masters degree.
The IIITH Experience
“I’m glad I came to IIITH. If I had to do it again, I would,” says the soft-spoken researcher when asked about his experience on campus. Marius is all praise for his mentor too, claiming that it was great to work with Prof. Varma. “Although we had briefly discussed probable research areas, it wasn’t until I got here and read papers and such for a month or so that I knew what I was going to work on. Basically we decided on a subject together. In a way, this was good because I ended up working on something really interesting to me,” he says. For Prof. Varma too, the experience was refreshing. “Unlike our students, international students are typically not worried about publications or a job, and hence his approach was a welcome change. Marius is technically very sound and was coming up with new ideas without worrying about publications. Many of our students lose out on the big picture because of pressures they face,” says Prof. Varma wistfully.
Indianisms
When quizzed about experiencing all things Hyderabad, Marius nods in the affirmative and remarks, “The amount of cars and people on the streets – that I have really experienced more than the monuments!” He politely says that he’s liked pretty much whatever he’s eaten so far, adding that while on campus he’s generally eaten only at Yuktahaar. “I have plain rice and not too spicy food”. Adopting a conservative approach, even outside of campus, he’s been mostly eating Indian food. “I was afraid to be a bit disappointed by Western food here, I mean.. expecting something and getting something different. I’ve tried biryani and haleem, some street food too, like dosa, and pani puri. Not a huge fan of it though,” he quips candidly. Prod him a little more about something that stands out from his Indian experience and he opens up about the great Indian head nod. ”People here tend to wobble their heads, which is to convey an ok. When you arrive with your visa at the airport, to interpret the ‘ok’ as a ‘no’ is a bit scary! This happened quite often when I had to deal with the administration. They look at your papers and you don’t know if it’s an ‘ok’ or if you’re in trouble!”.
In 5 months, Marius has done considerable research which is being carried forward by another Masters student from IIITH. The team is looking at publishing it soon at an upcoming conference. Further research is definitely Marius’ long-term vision. “I don’t really know where I want to do a Ph.D, but it will be in Machine Learning or Computer Vision though,” he said before bidding adieu.