[Paper Review] Score-Based Generative Models Detect Manifolds

Jakiw Pidstrigach|arXiv (Cornell University)|Jun 2, 2022

Generative Adversarial Networks and Image Synthesis4 citations

TL;DR

This paper establishes theoretical conditions under which score-based generative models (SGMs) can reliably generate samples from an underlying low-dimensional data manifold $ mathcal{M}$, even when approximating the true data distribution. It proves that SGMs can detect and generate from the correct manifold structure—such as faces or training data patterns—despite distributional inaccuracies or model approximations, providing a foundation for understanding generalization and memorization in diffusion models.

ABSTRACT

Score-based generative models (SGMs) need to approximate the scores $ abla \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.

Motivation & Objective

• To understand under what conditions score-based generative models (SGMs) can generate samples from the true data manifold $ mathcal{M}$, rather than just memorizing training data.
• To analyze the generalization properties of SGMs by characterizing when they memorize training data versus learn the underlying data manifold.
• To establish theoretical conditions under which SGMs robustly produce samples from the correct manifold, even when the final distribution $p_T$ and scores $\nabla\log p_t$ are approximated.
• To bridge the gap in theoretical understanding of how approximation errors in the prior and score network affect the final sample distribution $\mu_{\text{sample}}$.

Proposed method

• The paper models the forward and reverse SDEs of SGMs, with the forward SDE diffusing data into noise and the reverse SDE reconstructing samples from noise.
• It analyzes the KL divergence between the forward SDE's marginal $p_t$ and a standard Gaussian $n_t$, showing it tends to zero as $t \to \infty$ under mild moment conditions.
• The analysis uses De Bruijn's identity and conditional expectation to relate the score function $\nabla\log p_t$ to the $L^2$-projection of initial data onto noisy observations.
• It proves that if the data distribution $\mu_{\text{data}}$ is supported in a ball of radius $M$, then the KL divergence $KL(p_t \| n_t) \to 0$ as $t \to \infty$, ensuring the model recovers the correct manifold structure.
• The paper extends the result to general distributions via truncation and dominated convergence, showing convergence of the KL divergence even without compact support.
• It establishes that SGMs can detect the data manifold $\mathcal{M}$ when the score network and prior are sufficiently accurate, even if the final distribution is not perfectly matched.

Experimental results

Research questions

• RQ1 Under what conditions can score-based generative models produce samples that lie on the true data manifold $\mathcal{M}$, even if the distribution is not perfectly recovered?
• RQ2 How do approximation errors in the prior $\mu_{\text{prior}}$ and the score network $s_\theta(x,t)$ affect the model's ability to generate samples from the correct manifold?
• RQ3 When does an SGM memorize training data versus generalize to the underlying data manifold?
• RQ4 What theoretical guarantees exist for the reverse SDE to recover the correct manifold structure despite distributional approximations?
• RQ5 How does the KL divergence between the forward SDE's marginal and a standard Gaussian behave as time $t$ increases, and what does this imply for manifold detection?

Key findings

• The KL divergence $KL(p_t \| n_t)$ converges to zero as $t \to \infty$ when the data distribution has finite second moments, ensuring the forward process forgets its initial condition and the model can recover the correct manifold.
• If the data distribution is supported in a ball of radius $M$, the KL divergence $KL(p_t \| n_t) \to 0$ as $t \to \infty$, which implies that the reverse SDE can generate samples from the correct manifold structure.
• The paper proves that SGMs can detect the data manifold $\mathcal{M}$ even when the score function $\nabla\log p_t$ is approximated by a neural network $s_\theta(x,t)$, provided the approximation error is bounded.
• The analysis shows that SGMs can generate samples from the correct manifold even if the final distribution $p_T$ is approximated by a prior $\mu_{\text{prior}}$, as long as the forward process forgets its initial condition.
• The model generalizes to the underlying data manifold when the score network approximates the true score function well, and memorization of training data occurs only when the model fails to detect the manifold structure.
• The theoretical framework provides a precise condition under which SGMs memorize training data: when the model fails to detect the low-dimensional structure of the data manifold $\mathcal{M}$.