A generative adversarial network (GAN) parameterizes an approximation to a given class of images. In practice, GAN-based models describe real-world image classes much more effectively than conventional wavelet-based models. Image processing with a GAN amounts to solving a non-convex optimization problem, and first-order methods tend to perform well. We introduce a model for neural layers that is amenable to analysis and suggests the use of certain activation functions. Joint work with Soledad Villar (NYU).

This presentation is part of Minisymposium “MS21 - Recent mathematical advances in phase retrieval and computational imaging (2 parts)”
organized by: Mahdi Soltanolkotabi (University of Southern California) , Tamir Bendory (Princeton University) .

Authors:

Dustin Mixon (Department of Mathematics, Ohio State University)

Keywords:

deep learning, image reconstruction