45 min
Seminar, Hosted by Simons Institute for the Theory of Computing
High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
MIT
Recording
Abstract
Unbiased stochastic queries can support sampling whose cost grows only polylogarithmically with inverse target error. This talk applies the framework to diffusion generation with sufficiently accurate score estimates and to log-concave sampling with subexponential gradient noise. The latter result contrasts with stochastic convex optimization, where even simple additive noise can require polynomially many queries.
Topics
- Unbiased stochastic queries, search World Wide
- inverse target error, search World Wide
- diffusion generation, search World Wide
- score estimates, search World Wide
- log-concave sampling, broader topic: sampling
- subexponential gradient noise, broader topic: noise
- stochastic convex optimization, broader topic: optimization
- additive noise