Probing neural networks for science: What is it they are learning?  [slides]

Abstract

Neural Networks are finding increasing use in many areas of astronomy, but often act as "black boxes". Many techniques exist to probe in the internals of neural networks but not all are relevant to scientists. In this talk I discuss some of the techniques developed in computer vision to investigate what neural networks are learning, and discuss some of their benefits and problems when applied to astronomy. I introduce a simple technique to probe what neural networks have learned and apply it to networks trained to find strong gravitational lenses.