Distribution Matching Evolutionary Algorithms for Discovery

Abstract: A novel discovery is one which is both useful and surprising. For generative models, an output is useful if it has a high reward and is surprising if it has a low probability. In this talk I will detail a method for sampling high reward low probability outputs from generative models, outputs which we would colloquially call novel discoveries. This method is based on a framework interpreting evolutionary algorithms as implementing approximate probabilistic inference. I will conclude by sharing some potential extensions of this framework which would enable developing other, perhaps more performant, algorithms for discovery.

Bio: Yonatan is a DPhil student at Oxford working on probabilistic AI methods for automated discovery, with a special focus on developing principled evolutionary algorithms. Previously, Yonatan obtained his master’s in computer science from the University of Cambridge and worked on maps for autonomous vehicles at Mobileye. His PhD is funded by the AIMS CDT and a Rhodes scholarship.

Join Online - https://meet.google.com/vgs-cctr-esu