Institute for Advanced Study / Princeton University Joint Astrophysics Colloquium

Machine learning methods are having a huge impact in astrophysics. However, ML has a strong ontology — in which only the data exist — and a strong epistemology — in which a model is considered good if and only if it performs well on held-out training data. These philosophies are in strong conflict with our goals, practices, and philosophy in astrophysics. I identify locations in astrophysics practice at which the philosophy of ML is valuable (for example: in causal separations of foregrounds). I also show that there are contexts in which the use of ML methods introduce strong, unwanted, and un-correctable biases. The answer to the question posed in my title is “both.” Work in collaboration with Soledad Villar (JHU).