Introduction

Models that are tractable can be analytically evaluated and easily fit to data (e.g. a Gaussian or Laplace). However these models are unable to aptly describe structure in rich datasets. On the other hand, models that are flexible can be molded to fit structure in arbitrary data.

여기에서 제시하는 diffusion probabilistic model은 다음과 같은 특징을 지닌다.

extreme flexibility in model structure
exact sampling
easy multiplication with other distributions(e.g. posterior)
the model log likelihood and probability of individual states to be cheaply evaluated

Overview

… uses a Markov chain to gradually convert one distribution into another. … We build a generative Markov chain which converts a simple known distribution (e.g. a Gaus- sian) into a target (data) distribution using a diffusion process. we explicitly define the probabilistic model as the endpoint of the Markov chain. Since each step in the diffusion chain has an analytically evaluable probability, the full chain can also be analytically evaluated.

What is different point with others?

… using ideas from physics, quasi-static processes, and annealed importance sampling rather than from variational Bayesian methods.

이는 기존에 variational inference에서 ELBO를 가지고 최적화를 하던 것과 구분되는 부분인듯