Article by Jakiw Pidstrigach accepted for NeurIPS 2022

Congratulations! We are proud that Jakiw Pidstrigach's article "Score-Based Generative Models Detect Manifolds" (abstract below, arxiv link) has been accepted at NeurIPS 2022, one of the leading machine learning conferences.


"Score-based generative models (SGMs) need to approximate the scores ∇log pt of the intermediate distributions as well as the final distribution pT of the forward process. The theoretical underpinnings of the effects of these approximations are still lacking. We find precise conditions under which SGMs are able to produce samples from an underlying (low-dimensional) data manifold M. This assures us that SGMs are able to generate the "right kind of samples". For example, taking M to be the subset of images of faces, we find conditions under which the SGM robustly produces an image of a face, even though the relative frequencies of these images might not accurately represent the true data generating distribution. Moreover, this analysis is a first step towards understanding the generalization properties of SGMs: Taking M to be the set of all training samples, our results provide a precise description of when the SGM memorizes its training data."