I’ll be studying the effects of second-order statistics in natural data distributions on learning dynamics in vision models. We know from the no-free-lunch theorems that the success of any model/algorithm pair is predicated on a fortuitous alignment between the pair’s inductive bias and the structure of the data distribution. In vision, neuroscientists have long known that natural image distributions exhibit powerlaw statistics in Fourier space. I want to know: do vision models exploit this structure?