Talk

Imprecise probabilistic Machine Learning: Being Precise About Imprecision

  • Michele Caprio (University of Manchester)
E2 10 (Leon-Lichtenstein)

Abstract

In this talk, I will talk about the history of Imprecise Probabilities (IPs) from their inception in Philosophy, to their later adoption in Statistics and other sciences. I'll make the case for why IPs are useful and indeed needed in (Probabilistic) Machine Learning methodology and theory. I will conclude with a recent result in Imprecise Probabilistic Machine Learning theory concerning the ergodic behavior of Imprecise Markov Semigroups. Such a result allows us to study the long-term behavior of smooth input transitions for Convolutional Autoencoders, in the presence of uncertainty and ambiguity.