[Colloquium] Olivia Tsang’s MS Presentation/ May 8, 2024

[Devin Davis]

This is an announcement of Olivia Tsang’s MS Presentation
===============================================
Candidate:

Location: JCL 346

Time: Wednesday, May 8 at 12:00 PM
Title: Solving the Inverse Scattering Problem with Machine Learning-Based Iterative Refinement

Abstract:
Acoustic and electromagnetic waves can be used to probe materials in a non-destructive manner, such as in seismic imaging, radar detection, and medical imaging. However, it can be challenging to recover an image of the object from measurements of the outgoing wave. This is known as an inverse scattering problem, and it is difficult because the wave scattering process is in general nonlinear and non-local. When low-frequency probe waves are used, the scattering process is nearly linear but ignores features smaller than the diffraction limit (i.e., half a wavelength long). A well-established approach, known as “Recursive Linearization,” uses multi-frequency data to break the recovery process into a sequence of linear inverse problems, which enables stable and high-resolution reconstructions.
Inspired by the Recursive Linearization idea, we propose a machine learning-based method that explicitly breaks the reconstruction process into multiple refinement steps. Methods built on Recursive Linearization tend to be very computationally expensive, while existing machine learning-based methods tend to struggle in the strongly nonlinear scattering regime. We find that our method outperforms existing machine learning approaches, as well as new baselines that omit the iterative refinement formulation.
This work has been done jointly with Owen Melia, Rebecca Willett, Vasilis Charisopoulos, Jeremy Hoskins, and Yuehaw Khoo.

Advisor: Rebecca Willett

Committee: Rebecca Willett, Risi Kondor, and Yuehaw Khoo


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20240425/32a521e9/attachment-0001.html>