Skip to main content
Browse by:
GROUP

Learning to solve inverse problems

Event Image
Friday, February 25, 2022
3:30 pm - 4:30 pm
Yves Atchadé, Professor, Statistics, Boston University
Statistical Science Seminar Series

Inverse problems (IPs) are ubiquitous in science and engineering. Statistical solutions to IPs are classically formulation as posterior distributions over the quantities of interest or as MAP estimates. However these statistical solutions can be numerically costly to implement and become intractable in some applications (such as remote sensing) where very large numbers of IPs are solved. The talk is an introduction to an alternative approach to IPs that has developed recently in computational imaging which consists in learning directly an IP solution from data using deep neural networks and nonparametric regression methods. I will present some recent results that we have obtained on the posterior contraction of these models in the sparse regime. The enormous challenge of sampling from these posterior distributions will also be discussed.

This will be a Hybrid seminar. It will be held in 116 Old Chemistry at 3:30 - 4:30 pm and on Zoom.

https://duke.zoom.us/j/92397382385
Meeting ID: 923 9738 2385
Passcode: 425966

Type: LECTURE/TALK