On Characterizing the Capacity of Neural Networks using Algebraic Topology

William H. Guss, Ruslan Salakhutdinov

The learnability of different neural architectures can be characterized directly by computable measures of data complexity. In this paper, we reframe the problem of architecture se…