Abstract: Inverse problems are everywhere — from imaging and geophysics to medicine and engineering — but solving them often means wrestling with expensive models and elusive uncertainties.
In this talk, we explore a new perspective: using paired autoencoders to learn the structure of inverse problems directly from data. This approach bypasses the need for forward model evaluations during inference, offering a powerful, flexible and fast alternative to traditional methods. By jointly learning compact representations of both data and solutions, we unlock new possibilities for robust estimation, real-time applications and generalization across problem domains. This is a step toward rethinking how we solve inverse problems in the era of machine learning.
Bio: Matthias Chung is an associate professor of Mathematics at Emory University.
See all upcoming talks at https://www.anl.gov/mcs/lans-seminars.