We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Score-Based Generative Models Detect Manifolds

Abstract: Score-based generative models (SGMs) need to approximate the scores $\nabla \log p_t$ of the intermediate distributions as well as the final distribution $p_T$ 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 $\mathcal{M}$. This assures us that SGMs are able to generate the "right kind of samples". For example, taking $\mathcal{M}$ 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 $\mathcal{M}$ to be the set of all training samples, our results provide a precise description of when the SGM memorizes its training data.
Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG); Numerical Analysis (math.NA); Probability (math.PR)
MSC classes: 68T99
ACM classes: I.2.0
Cite as: arXiv:2206.01018 [stat.ML]
  (or arXiv:2206.01018v3 [stat.ML] for this version)

Submission history

From: Jakiw Pidstrigach [view email]
[v1] Thu, 2 Jun 2022 12:29:10 GMT (4308kb,D)
[v2] Wed, 17 Aug 2022 09:18:51 GMT (12119kb,D)
[v3] Sat, 15 Oct 2022 12:54:18 GMT (12119kb,D)

Link back to: arXiv, form interface, contact.