Title: Memory and Energy: Two Bottlenecks for Learning
Abstract: For many large-scale learning systems, two critical bottlenecks are 1) the amount of memory required, and 2) the amount of energy required both in training and at inference time. (In certain settings, the role of these resources may now eclipse the more traditional resources of time, and the amount of training data.) In this talk, I’ll discuss the current landscape of memory bounded learning and optimization, and will present some new perspectives on energy-centric models of computing and learning, together with a proposal for a low-energy learning system. Throughout, the emphasis will be on the many open directions on both fronts.
Bio: Gregory Valiant is an Associate Professor at Stanford, where he works on a wide variety of problems across learning, information theory, algorithms, and optimization.