48 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Finite-particles rates for drifting models
UC Davis
Recording
Abstract
A Wasserstein-gradient-flow perspective suggests a conservative alternative for one-step generative models. This talk replaces a displacement velocity with the difference of smoothed data and model scores, derives finite-particle bounds through a joint-entropy identity, and compares the resulting guarantees with those for a non-conservative Laplace-kernel method.
Topics
- Wasserstein-gradient-flow, broader topic: flow
- one-step generative models, broader topic: generative models
- displacement velocity, search World Wide
- model scores, search World Wide
- finite-particle bounds, search World Wide
- joint-entropy identity, broader topic: identity
- non-conservative Laplace-kernel method, search World Wide
- conservative alternative, search World Wide