Hosted by Peter McMahon

Title: Neural Network Priors for Inverse Problems with Less Data

Abstract: Inverse problems where one tries to understand the cause of observed effects by “inverting” a forward transformation are ubiquitous in science. Leveraging prior information about the structure of the underlying causes greatly improve the recovery. With the recent successes of unsupervised learning, neural network priors are able to encode very rich structure of signals. We will see how they can help to solve inverse problems with little data, discuss their strengths but also weaknesses. Finally, we will see how they can be useful even in the absence of training data for imaging inverse problems. As an example, we will see how they can help overcome challenging conditions in a concrete experimental setting.

This talk is based on joint works with Eric Tramel, Francesco Caltagirone, Florent Krzakala, Hannah Lawrence, David Bramherzig, Henri Li and Michael Eickenberg.

Short bio: I completed my PhD in Statistical Physics in 2019 under the supervision of Florent Krzakala and Lenka Zdeborová at École Normale Supérieure (Paris). For my thesis, I worked on the statistical mechanics of learning: using computation methods from the physics of disordered systems to understand learning in neural networks. Since January 2020, I am a postdoc co-affiliated between NYU center for data science and Flatiron Institute. Besides theoretical machine learning, I am also interested in applications of machine learning to scientific discoveries (which is more the topic of this talk).