VC dimension of neural nets (Alexander Kozachinskiy)
Consider a neural network with W parameters and with sign as the activation function. How large can its VC dimension be? I will explain a folklore upper bound of O(W log W), and will give an example of Maass 1994, establishing that this bound is tight even for constant-depth neural networks.