Kyle Cranmer - New York University (NYU) The Physical Sciences are an interesting environment for deep learning techniques primarily because so much is known about the generative model for the data. Some fields have detailed simulations that can be incorporated directly into approximate inference techniques or used to generate labeled training data for supervised learning. Often domain knowledge motivates hierarchical or causal structures that can be leveraged in deep learning models. At the same time, the demands of scientific problems often motivate new training objectives or challenge common assumptions. I will give examples of these observations.