Abstract

In many scientific and medical settings, we cannot directly observe images of interest, such as a person’s internal organs, the microscopic structure of materials or cells, or distant stars and galaxies. Rather, we use MRI scanners, microscopes, and satellites to collect indirect data that require sophisticated algorithms to form an image. Historically, these methods have relied on mathematical models of simple image structures to improve the quality and resolution of the resulting images. In this talk, I will describe recent efforts to harness vast collections of images to train computers to learn more complex models of image structure, yielding more accurate and higher-resolution images than ever. These new methods lead to new insights into designing neural networks in a principled manner to jointly leverage both training data and physical models of how imaging data is collected. We will conclude with a discussion of some of the main open questions and exciting new directions in this emerging field.