Description:

The SkAI Institute will host a colloquium featuring Berthy Feng, Postdoctoral Fellow at Schwarzman College of Computing at the Massachusetts Institute of Technology as well as at the NSF AI Institute for Artificial Intelligence and Fundamental Interactions.

Coffee and pastries will be provided prior to the event. Lunch will be provided afterwards.

Talk title: “Taming Priors for Scientific Computational Imaging”

Abstract: Computational imaging aims to visualize scientific phenomena beyond the reach of conventional optics by incorporating assumptions, or priors, about the object being imaged. In the age of AI, the priors available to us are more sophisticated than ever. However, it is yet unclear how to use these priors rigorously, especially when using imaging in the scientific process. The main challenge of computational imaging for science is discerning when to use what prior and how. The art of imaging lies in deciding the correct balance of assumptions to obtain trustworthy and informative images. Once the prior has been decided, the question is how to incorporate it rigorously.

In this talk, I will present my previous work on building principled routes for incorporating data-driven and physics-based priors. On the data-driven side, I will show results of re-imagining the famous M87 black hole from real data with score-based priors. On the physics-based side, I will show how we have tackled extremely under-determined imaging problems by enforcing physics constraints, including the problem of single-viewpoint dynamic tomography of emission near a black hole. In the intersection of AI and physics, I will present neural approximate mirror maps, a way to enforce physics constraints on generative models. I will then discuss future directions for further taming priors so that we can rigorously create, interpret, and extract insights from scientific images.

Register by Monday, February 16 at 12:00 p.m. at the link below.