Title: From differential learning to diffusion models
Authors: Samuel Kaski - The University of Manchester (United Kingdom) [presenting]
Markus Heinonen - Aalto University (Finland)
Abstract: Deep learning models construct successive representations across layers, and many DL models have recently been shown to converge to various stochastic processes (such as SDEs). This new perspective has fueled dramatic developments in machine learning, where generative models such as DALLE or Imagen exhibit jaw-dropping performance. We will survey some of the foundational principles of such models, and demonstrate how such principles can be used to re-think supervised learning as well.