Current browse context:
cond-mat
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: Self-Consistent Dynamical Field Theory of Kernel Evolution in Wide Neural Networks
(Submitted on 19 May 2022 (v1), last revised 4 Oct 2022 (this version, v3))
Abstract: We analyze feature learning in infinite-width neural networks trained with gradient flow through a self-consistent dynamical field theory. We construct a collection of deterministic dynamical order parameters which are inner-product kernels for hidden unit activations and gradients in each layer at pairs of time points, providing a reduced description of network activity through training. These kernel order parameters collectively define the hidden layer activation distribution, the evolution of the neural tangent kernel, and consequently output predictions. We show that the field theory derivation recovers the recursive stochastic process of infinite-width feature learning networks obtained from Yang and Hu (2021) with Tensor Programs . For deep linear networks, these kernels satisfy a set of algebraic matrix equations. For nonlinear networks, we provide an alternating sampling procedure to self-consistently solve for the kernel order parameters. We provide comparisons of the self-consistent solution to various approximation schemes including the static NTK approximation, gradient independence assumption, and leading order perturbation theory, showing that each of these approximations can break down in regimes where general self-consistent solutions still provide an accurate description. Lastly, we provide experiments in more realistic settings which demonstrate that the loss and kernel dynamics of CNNs at fixed feature learning strength is preserved across different widths on a CIFAR classification task.
Submission history
From: Blake Bordelon [view email][v1] Thu, 19 May 2022 16:10:10 GMT (5588kb,D)
[v2] Mon, 26 Sep 2022 13:09:26 GMT (5670kb,D)
[v3] Tue, 4 Oct 2022 15:55:59 GMT (5672kb,D)
Link back to: arXiv, form interface, contact.