CNO - The power to efficiently approximate causal operators between infinite-dimensional linear spaces

Anastasis Kratsios, McMaster University

11/28, 2022 at 11:30am-12:30pm (HH403)

In this talk, we overview recent developments in the causal approximation of stochastic phenomena. We introduce the causal neural operator, a principled deep neural model design framework which, given two infinite-dimensional linear metric spaces X and Y, builds a universal deep learning model mapping sequences of inputs in X to sequences of outputs in Y, causally. As an application, we generically solve the robust stochastic filtering problem. Our proof also produces a non-recursive, federated, and parallelizable learning algorithm for training the CNO. I'll also discuss potential applications to PDE and Besov spaces. Project marketing will then follow.