Theoretical Analysis of the Inductive Biases in Deep Convolutional Networks