Parsimony and Overfitting

Abstract Philosophers often defend appeals to parsimony by invoking its central role in science. I argue that this move fails once we distinguish between two uses of parsimony: non-ideal and ideal . Non-ideal parsimony enjoys strong inductive support in science, since complex models are prone to overfit to predictively irrelevant noise. But philosophical data aren’t significantly noisy in the relevant sense: when our intuitions are unreliable, their unreliability typically reflects systematic bias rather than noise, which parsimony doesn’t mitigate. Philosophers therefore need ideal parsimony, which finds only weak support from science. Thus, the scientific analogy cannot vindicate the philosopher’s use of parsimony.