50 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Understanding and enhancing diffusion model: a quantification of its generalizability, and a provably scalable test-time scaling method
Georgia Tech
Recording
Abstract
The first part characterizes how continuous diffusion models generate beyond their training examples using the empirical distribution itself, clarifying their inductive bias without a population-limit assumption. The second introduces reward-guided remasking for discrete generation without changing model weights. The proposed sampler corrects earlier choices and comes with robustness guarantees under noisy process verification, contrasting its quadratic complexity with potentially exponential error accumulation in common selection heuristics.
Topics
- continuous diffusion models, broader topic: diffusion models
- empirical distribution, search World Wide
- inductive bias
- population-limit assumption, search World Wide
- reward-guided remasking, search World Wide
- discrete generation, search World Wide
- model weights
- proposed sampler, search World Wide
- noisy process verification, broader topic: verification
- quadratic complexity, broader topic: Complexity