Special Data Science Seminar - Anastasis Kratsios - Universal Stochastic Processes


Speaker: Anastasis Kratsios - University of Basel

Title: Universal Stochastic Processes

Abstract: Regular conditional distributions (RCDs) are of central importance to most areas of applied probability, ranging from mathematical finance to uncertainty quantification in machine learning. Nevertheless, there is currently no available universal class of deep neural models capable of approximating any RCD. We fill this theoretical gap by proposing a new geometric deep learning model with inputs in a Euclidean and out puts in the Adapted Wasserstein space over a Euclidean space across any finite time horizon. We then use our universal "static'' deep learning models to build a recursive neural network model which can approximate any stochastic process across an infinite time horizon. Numerical experiments illustrate our model's effectiveness.

Joint work: with Beatrice Acciaio and Gudmund Pammer.

Virtual Meeting - Zoom Meeting

Meeting ID: 949 5373 0428
Passcode: 126157
Go Back
McMaster University - Faculty of Science | Math & Stats