AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-20
Gaussian equivalence for CNN with non-Gaussian weights
This paper studies the infinite-width limit of convolutional neural networks with random weights. It shows that the Gaussian process limit holds without Gaussian priors, provided the weights are i.i.d., centered, and have enough finite moments. The framework covers polynomial-growth activations and common patch-extraction operators, including convolution, padding, stride, and pooling.