Theoretical Foundations of Score-Based Modeling and a Case Study on Stochastic Localization


Loading...

Date

2025

Publication Type

Student Paper

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

Diffusion models have emerged as a powerful framework for generative modeling, achieving state-of-the-art results in tasks such as image synthesis and density estimation. These models are typically formulated in a probabilistic setting, where the goal is to learn a generative process that, given i.i.d. samples from an unknown target distribution μ, can generate new samples from the same distribution. Diffusion models achieve this by learning the score function, i.e., the gradient of the log-likelihood. After briefly comparing diffusion models with other approaches used in generative modeling, we introduce methods for learning the score function, such as score matching [3] and denoising score matching [4]. Once the score function is learned, we explore various sampling algorithms, including annealed Langevin dynamics [6], denoising diffusion probabilistic models (DDPM) [7], score-based generative modeling via SDEs [8], in particular with the OU process, and stochastic localization [9]. We then analyze key results from the paper by Montanari [9], which provides further insights into the empirical properties of these methods. In the end, following the approach of Shah et al. [10], we study the convergence properties of stochastic localization when applied to a two-component Gaussian mixture model. All code used to produce the results and figures in this work is available at the author's Github.

Publication status

published

External links

Editor

Contributors

Examiner : Chen, Yuansi

Book title

Journal / series

Volume

Pages / Article No.

Publisher

ETH Zurich

Event

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Score-Based Generative Modeling, Diffusion Models, Denoising Score Matching, Stochastic Localization, Gaussian Mixture Models

Organisational unit

09819 - Chen, Yuansi / Chen, Yuansi check_circle

Notes

Funding

Related publications and datasets