Depth Separation in Norm-Bounded Infinite-Width Neural Networks

We study depth separation in infinite-width neural networks, where complexity is controlled by the overall squared $\ell_2$-norm of the weights (sum of squares of all weights in the network). Whereas previous depth separation results focused on separation in terms of width, such results do not give insight into whether depth determines if it is possible to learn a network that generalizes well even when the network width is unbounded. Here, we study separation in terms of the sample complexity required for learnability. Specifically, we show that there are functions that are learnable with sample complexity polynomial in the input dimension by norm-controlled depth-3 ReLU networks, yet are not learnable with sub-exponential sample complexity by norm-controlled depth-2 ReLU networks (with any value for the norm). We also show that a similar statement in the reverse direction is not possible: any function learnable with polynomial sample complexity by a norm-controlled depth-2 ReLU network with infinite width is also learnable with polynomial sample complexity by a norm-controlled depth-3 ReLU network.

Linear Neural Network Layers Promote Learning Single- and Multiple-Index Models

This paper explores the implicit bias of overparameterized neural networks of depth greater than two layers. Our framework considers a family of networks of varying depths that all have the same capacity but different implicitly defined representation costs. The representation cost of a function induced by a neural network architecture is the minimum sum of squared weights needed for the network to represent the function; it reflects the function space bias associated with the architecture. Our results show that adding linear layers to a ReLU network yields a representation cost that favors functions that can be approximated by a low-rank linear operator composed with a function with low representation cost using a two-layer network. Specifically, using a neural network to fit training data with minimum representation cost yields an interpolating function that is nearly constant in directions orthogonal to a low-dimensional subspace. This means that the learned network will approximately be a single- or multiple-index model. Our experiments show that when this active subspace structure exists in the data, adding linear layers can improve generalization and result in a network that is well-aligned with the true active subspace.