Analyzing Inverse Problems with Invertible Neural Networks

  • 2018-08-14 14:58:59
  • Lynton Ardizzone, Jakob Kruse, Sebastian Wirkert, Daniel Rahner, Eric W. Pellegrini, Ralf S. Klessen, Lena Maier-Hein, Carsten Rother, Ullrich Köthe
  • 12

Abstract

In many tasks, in particular in natural science, the goal is to determinehidden system parameters from a set of measurements. Often, the forward processfrom parameter- to measurement-space is a well-defined function, whereas theinverse problem is ambiguous: one measurement may map to multiple differentsets of parameters. In this setting, the posterior parameter distribution,conditioned on an input measurement, has to be determined. We argue that aparticular class of neural networks is well suited for this task -- so-calledInvertible Neural Networks (INNs). Although INNs are not new, they have, sofar, received little attention in literature. While classical neural networksattempt to solve the ambiguous inverse problem directly, INNs are able to learnit jointly with the well-defined forward process, using additional latentoutput variables to capture the information otherwise lost. Given a specificmeasurement and sampled latent variables, the inverse pass of the INN providesa full distribution over parameter space. We verify experimentally, onartificial data and real-world problems from astrophysics and medicine, thatINNs are a powerful analysis tool to find multi-modalities in parameter space,to uncover parameter correlations, and to identify unrecoverable parameters.

 

Quick Read (beta)

loading the full paper ...