The next D5 Seminar, “Bio-inspired self-supervised learning of visual representations” will be held by Arthur Aubret, on Friday, December 13 at 1:30 p.m. in room C005.
Abstract: Despite recent advances in self-supervised visual machine learning, humans develop more robust representations with much fewer data. This may be explained by the fundamental differences between the development of their visual systems: while machine learning methods use massive amounts of i.i.d images, humans actively move and interact with objects over time. In this talk, I investigate how considering bio-inspired learning mechanisms can impact visual representations learning. I will provide evidence that modelling spatio-temporal regularities in egocentric visual sequences boosts the robustness of vision models. In addition, I will explain how egocentric actions underpinning visual changes, like eye saccades or object manipulations, support object learning. Together, these findings expose that the spatio-temporal structure and active nature of human visual experience may be key to develop strong semantic visual representations.
