Learning to Act More Like Humans, and Learning with Less Data
Humans recognize objects, read and understand language, observe the world and control their bodies with incredible skill. Furthermore, we learn to do these things with very little training data. In this talk I survey recent work from my group and collaborators on deep learning techniques for these settings. I'll focus on the theme of learning with less labelled data using techniques such as multi-task learning, encoder-decoder architectures and a new technique for visual imitation learning with reinforcement learning (VIRL). I'll show how this new VIRL method permits a humanoid robot living in a software simulation to learn new behaviours by simply watching video of a desired activity.
Christopher Pal is a Full Professor in the department of information and software engineering at Polytechnique Montreal, one of the founding faculty members of Mila, the Quebec AI institute, and a Principal Research Scientist at Element AI. He has served as an Area Chair for conferences such as CVPR, ICCV, NeurIPS, ICML and ICLR and he will be one of the program chairs for ICCV 2021 in Montreal. He is also one of the general chairs of the conference MIDL 2020. He also holds a Canada CIFAR AI chair.