Abstract:Denoising diffusion models have become ubiquitous for generative modeling. The core idea is to transport the data distribution to a Gaussian by using a diffusion. Approximate samples from the data distribution are then obtained by estimating the time-reversal of this diffusion using score matching ideas. We follow here a similar strategy to sample from unnormalized probability densities and compute their normalizing constants. However, the time-reversed diffusion is here simulated by using an original iterative particle scheme relying on a novel score matching loss. Contrary to standard denoising diffusion models, the resulting Particle Denoising Diffusion Sampler (PDDS) provides asymptotically consistent estimates under mild assumptions. We demonstrate PDDS on multimodal and high dimensional sampling tasks.

What problem does this paper attempt to address?

This paper mainly discusses how to use Particle Denoising Diffusion Sampler (PDDS) to solve the problem of Monte Carlo sampling and estimation of normalization constant from an unnormalized probability density. Traditional denoising diffusion models are commonly used for generative modeling, but in this paper, the authors propose a new approach that applies this idea to sampling tasks. The core idea of the paper is to gradually transform the target distribution into a Gaussian distribution through the noise diffusion process, and then sample from the target distribution through the reverse diffusion process in a theoretical manner. However, due to the inability to simulate this process exactly, especially due to the infeasibility of gradient terms, the authors propose a new strategy based on iterative particle methods. This method relies on a new score matching loss to provide asymptotically consistent estimation under mild assumptions. The main contributions of the paper include: 1. Adapting guided denoising diffusion models to sampling problems; 2. Providing theoretical results to quantify the error of the current guided denoising diffusion models under simple scenarios; 3. Developing a Sequential Monte Carlo (SMC) scheme to provide consistent estimation in this setting and establishing limit theorems; 4. Introducing an algorithm to reduce the variance of SMC estimates based on the new score matching loss. The authors also point out that the main source of error in existing methods is the inability to close the guided term expression, which leads to a decrease in the sampling quality in high-dimensional cases. They correct this issue through particle methods (also known as sequential Monte Carlo), which can provide consistent estimation as the number of particles tends to infinity. In addition, the paper proposes a neural network parameterization strategy that uses score matching to learn latent functions and further improves these estimates through iterative training. Finally, the experimental section demonstrates the performance of PDDS in estimating normalization constants in various sampling tasks, comparing it with existing methods such as SMC, CRAFT (Variational Autoencoder), DDS (Differential Diffusion Sampler), and PIS (Importance Sampling). Overall, the goal of the paper is to address the problem of effective sampling from complex distributions without samples. It provides a sampling framework with advantages in both theory and practice through improved particle methods and score matching techniques.

Particle Denoising Diffusion Sampler

Denoising Diffusion Samplers

Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models

From Denoising Diffusions to Denoising Markov Models

Denoising Diffusion Step-aware Models

Filling Deterministic Approximate Gap in Denoising Diffusion Probabilistic Models with Energy

Semi-Implicit Denoising Diffusion Models (SIDDMs)

Learning to Discretize Denoising Diffusion ODEs

Directly Denoising Diffusion Models

UDPM: Upsampling Diffusion Probabilistic Models

Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions

Mix-DDPM: Enhancing Diffusion Models Through Fitting Mixture Noise with Global Stochastic Offset

Diffusion-PINN Sampler

Interpreting and Improving Diffusion Models from an Optimization Perspective

Denoising Diffusion Bridge Models

To smooth a cloud or to pin it down: Guarantees and Insights on Score Matching in Denoising Diffusion Models

A Geometric Perspective on Diffusion Models

Listening to the Noise: Blind Denoising with Gibbs Diffusion

Divide-and-Conquer Posterior Sampling for Denoising Diffusion Priors

Pseudo Numerical Methods for Diffusion Models on Manifolds