Langevin Dynamics, Deep Learning and Bayesian Inference

Speaker: Natan Katz Abstract: Langevin Dynamics (LD) is an exciting mathematical tool for ML world, though it's not yet widely known. It leverages ideas from differential equations, stochastic processes, and 20th-century numerics. This talk will show how LD improves the interface between Deep Learning and Bayesian Inference. We will also discuss how Stochastic Gradient LD (SGLD) can help with Bayesian neural networks, and present a cool PyTorch implementation. Bio: Natan Katz has a BSc in Math and Physics from the Hebrew University, and a MSc in Applied Math from Weizmann Institute with focus on nonlinear dynamics. He spent one year in Frankfurt University. He has worked in biometrics, NLP, quantitative analysis, and more. In the past year, he has worked in cybersecurity, first at Avanan, now at Checkpoint. He also writes for TDS ().