MS Teams
Speaker
Bamdad Hosseini | California Institute of Technology
Title
Inverse Problems and Machine Learning
Abstract
The fields of inverse problems and machine learning have a lot in common. In fact, many of the algorithms and problems in each of these fields can be analyzed and developed from the perspective of the other. In this talk I give an overview of some of my research at the intersection of these fields. In the first part I will discuss the asymptotic consistency of a graphical semi-supervised learning algorithm. Using ideas from theory of inverse problems and spectral analysis of elliptic operators to develop a deeper understanding of how and why graphical algorithms work. In the second part of the talk, I will present an algorithm for data-driven solution of Bayesian inverse problems by combining tools from machine learning, such as generative adversarial networks, with ideas in measure transport. This approach leads to a model agnostic method for conditional sampling and in turn the solution of inverse problems.