Abstract: While deep learning has been described as a largely empirical discipline, where practice leads theory, many of the approaches of machine learning research don’t appear to be grounded in the rigorous tradition of the empirical sciences. To understand the inner workings of deep networks at a fundamental level, these models ought to be studied like physical objects that follow specific dynamics governed by laws of motion, and examined with the tools of the scientific method, to ensure we not only understand effect, but also begin to understand cause, which is the raison d'être of science. In the context of AI fairness and pruning, I’ll show examples of how hypothesis testing, reproducible workflows, and careful experiment design can help validate or refute hypotheses.