Princeton University Thunch Talk

I will present a review of how deep learning is used in astrophysics, and how this use is often misguided. I will introduce the term “scientific debt,” and argue that, though deep learning can quickly solve a complex problem, its success does not come for free. Because of the fact that most learning techniques are completely opaque in their conclusions, we as a community accumulate scientific debt whenever machine learning is used without an interpretation. Fundamentally, science is about understanding, and when we blindly learn and exploit a pattern in a high-dimensional dataset, we have not gained any new scientific insight; we have avoided the underlying problem and taken on debt, left to some future scientist to repay.

In the second part of the talk, I will outline several methods created by myself and collaborators over the last year to help address these problems which use symbolic regression. I will show how one can derive interpretable analytic relations from trained deep neural networks; this allows us to extract insight from machine learning used in astrophysics. I will conclude by showing several applications to astrophysics by us and others who have implemented our techniques, and how we may gain new insights from such results.