A1288
Title: Recent advances in understanding efficiencies of diffusion models: Accelerated convergence and statistical optimality
Authors: Gen Li - The Chinese University of Hong Kong (Hong Kong) [presenting]
Changxiao Cai - University of Michigan (United States)
Abstract: Score-based diffusion models have become a foundational paradigm for modern generative modeling. Some recent progress will be presented towards understanding and improving the statistical and computational efficiencies of diffusion models. The first aim focuses on computational efficiency, where we propose a novel acceleration sampling scheme for stochastic samplers that provably improves the iteration complexity under minimal assumptions. The second work is concerned with the statistical efficiency. We develop the first end-to-end theoretical framework for deterministic samplers and establish the (near)-minimax optimal guarantees under mild assumptions on target data distributions.
