Why Deep Regression?

Hi, Megh here ,

I am a Machine Learning and Computer Vision Ph.D. student at EPFL, supervised by Prof. Alexandre Alahi, Visual Intelligence for Transportation Lab and Dr. Mathieu Salzmann . My thesis is centred around different methodologies and applications of deep hetereoscedastic regression. One may ask, why deep regression? My reasoning is that regression is ubiquitous in machine learning. I believe that a better understanding of regression will lead to a better understanding of machine learning as a whole. Why heteroscedastic? Heteroscedasticity is inherent in our world! For instance, the uncertainty in the price of a house is not the same as … let’s say … a loaf of bread! The former is more uncertain than the latter. In fact, the price of both a house and a loaf of bread is a function of where in the world we are in!

With this thesis, I wish to contribute to research in deep regression. To this end, we propose different methologies, aimed at fundamentally improving deep heteroscedastic regression as well as evaluation metrics to measure its accuracy! I am especially proud of these works since in my opinion they provide fundamental advancements in machine learning, and will remain relevant for years to come (hopefully!). In this thesis, we also study different applications of deep regression. One such application is in human pose forecasting, a challenging problem due to the inherent ambiguity in the different future motions a person can undertake! I am now interested in the study of deep regression in generative AI, which will also be the culmination of my thesis.

Please check out my research, and as usual, feedback is welcome! Also, the link to my CV for reference. My personal homepage lists my different research experience at IIT Bombay, Mercedes-Benz as well as random thoughts on different topics!