Mean Field Residual Networks: On the Edge of Chaos

TL;DR

This paper analyzes residual networks using mean field theory, revealing how skip connections lead to subexponential and polynomial dynamics that preserve input geometry and gradient flow, and predicts optimal initializations for training.

Contribution

It introduces a theoretical framework for understanding residual networks' dynamics, deriving polynomial exponents, and predicting optimal initializations based on network depth.

Findings

Residual networks exhibit subexponential and polynomial dynamics depending on nonlinearity.

Theoretical predictions accurately match empirical training performance on MNIST.

Common initializations like Xavier or He are suboptimal for residual networks, with optimal variances depending on depth.

Abstract

We study randomly initialized residual networks using mean field theory and the theory of difference equations. Classical feedforward neural networks, such as those with tanh activations, exhibit exponential behavior on the average when propagating inputs forward or gradients backward. The exponential forward dynamics causes rapid collapsing of the input space geometry, while the exponential backward dynamics causes drastic vanishing or exploding gradients. We show, in contrast, that by adding skip connections, the network will, depending on the nonlinearity, adopt subexponential forward and backward dynamics, and in many cases in fact polynomial. The exponents of these polynomials are obtained through analytic methods and proved and verified empirically to be correct. In terms of the "edge of chaos" hypothesis, these subexponential and polynomial laws allow residual networks to "hover over the boundary between stability and chaos," thus preserving the geometry of the input space and the gradient information flow. In our experiments, for each activation function we study here, we initialize residual networks with different hyperparameters and train them on MNIST. Remarkably, our initialization time theory can accurately predict test time performance of these networks, by tracking either the expected amount of gradient explosion or the expected squared distance between the images of two input vectors. Importantly, we show, theoretically as well as empirically, that common initializations such as the Xavier or the He schemes are not optimal for residual networks, because the optimal initialization variances depend on the depth. Finally, we have made mathematical contributions by deriving several new identities for the kernels of powers of ReLU functions by relating them to the zeroth Bessel function of the second kind.