Towards understanding theoretical limitations of meta learnersSpeaker: James Lucas, University of TorontoTime: Tuesday, June 15, 2021, 10:00AM - 11:00AM, EST Zoom Link: contact tml.online.seminars@gmail.com Abstract:Machine learning models have traditionally been developed under the assumption that the training and test distributions match exactly. However, recent success in few-shot learning and related problems are encouraging signs that these models can be adapted to more realistic settings where train and test distributions differ. In this talk, I will present novel information-theoretic lower-bounds on the error of learners that are trained on data from multiple sources and tested on data from unseen distributions. These bounds depend intuitively on the information shared between sources of data, and characterize the difficulty of learning in this setting for arbitrary algorithms (in a minimax sense). I will conclude with an application of these bounds to a Hierarchical Bayesian model of meta learning, providing additional risk upper bounds in this setting.Speaker's BioJames Lucas is a PhD candidate at the University of Toronto, where he is supervised by Richard Zemel and Roger Grosse. James’ research has primarily focused on the practice and theory behind training deep neural networks. In particular, he has developed techniques for training provably smooth neural networks and investigated the loss landscape geometry of deep nets. More recently, he has studied theoretical limitations of learning in realistic settings such as few-shot learning. In addition to his research, James is passionate about video game development and is currently working towards releasing his first game. |