Core Concepts
Wide neural networks can be accelerated by modifying the kernel spectrum to control the inductive bias efficiently.
Abstract
Introduction to the bias of wide neural networks towards learning specific functions.
The relationship between network width and convergence to a kernel regression solution with NTK.
Characterization of spectral bias in wide neural networks based on NTK decomposition.
Proposal of Modified Spectrum Kernels (MSKs) for approximating kernels efficiently.
Introduction of preconditioned gradient descent method for accelerating network training.
Demonstration that accelerated convergence does not alter final predictions.
Algorithmic details on constructing MSKs and preconditioning for spectral bias manipulation.
Experiments validating the efficiency and effectiveness of the proposed method.
Stats
arXiv:2307.14531v2 [cs.LG] 20 Mar 2024