Skip to main content

Ankush Gupta Seminar

Friday 3rd December @12 noon

Ankush Gupta (DeepMind/AIMS Alumni) - Self-supervised Visual Representation Learning for Perception and Control

Abstract: This talk aims at describing methods for learning spatially grounded object-centric visual keypoint detectors in a self-supervised manner, and their use in downstream control tasks. While supervised learning is the workhorse of modern deep learning methods, and self-supervised learning is primarily aimed at learning abstract feature representations, few methods have been developed which can directly (without any labelled data) output spatially grounded representations. The evolution of self-supervised keypoint learning methods is traced through a discussion of three distinct methods, and their use is demonstrated in control domains, where keypoints enable sample-efficient reinforcement learning, and deep exploration due to a drastically reduced search space. Finally, keypoint-like representations are used for spatially-aware matching to achieve highly-performant few-shot classification across diverse visual domains.

Bio: Ankush Gupta is a research scientist at DeepMind, London. His research interests include developing models for learning flexible and transferable representations of visual data (images/videos). He obtained his DPhil (Computer Vision) in 2018 working in the Visual Geometry Group (VGG), Oxford with Prof. Zisserman and Prof. Vedaldi, and bachelors degree in 2014 in Electrical Engineering and Computer Science, from the University of California, Berkeley. He was part of the AIMS CDT programme while at Oxford.