English

Universal Smoothed Score Functions for Generative Modeling

Machine Learning 2023-03-22 v1 Machine Learning

Abstract

We consider the problem of generative modeling based on smoothing an unknown density of interest in Rd\mathbb{R}^d using factorial kernels with MM independent Gaussian channels with equal noise levels introduced by Saremi and Srivastava (2022). First, we fully characterize the time complexity of learning the resulting smoothed density in RMd\mathbb{R}^{Md}, called M-density, by deriving a universal form for its parametrization in which the score function is by construction permutation equivariant. Next, we study the time complexity of sampling an M-density by analyzing its condition number for Gaussian distributions. This spectral analysis gives a geometric insight on the "shape" of M-densities as one increases MM. Finally, we present results on the sample quality in this class of generative models on the CIFAR-10 dataset where we report Fr\'echet inception distances (14.15), notably obtained with a single noise level on long-run fast-mixing MCMC chains.

Keywords

Cite

@article{arxiv.2303.11669,
  title  = {Universal Smoothed Score Functions for Generative Modeling},
  author = {Saeed Saremi and Rupesh Kumar Srivastava and Francis Bach},
  journal= {arXiv preprint arXiv:2303.11669},
  year   = {2023}
}

Comments

Technical Report

R2 v1 2026-06-28T09:25:45.815Z