Pruned Neural Networks are Surprisingly Modular

Daniel Filan, Shlomi Hod, Cody Wild +2

The learned weights of a neural network are often considered devoid of scrutable internal structure. To discern structure in these weights, we introduce a measurable notion of modu…