Machine learning is presenting new opportunities for astrophysics—and all the natural sciences. The standard machine-learning workflow represents a very different epistemology than that of other kinds of scientific methods. How does that impact our results and beliefs about those results? I’ll discuss the different roles for machine learning in astrophysics and discuss them in terms of their effects on measurement precision and understanding. I’ll argue that it’s very different to, say, use a Gaussian Process to model stellar variability in a transit measurement than it is to, say, replace an n-body simulation with a deep-learning emulator.
