FLAIR · project
PIS: Path Integral Sampler: A Stochastic Control Approach for Sampling
Many sampling problems specify a probability density only up to an unknown normalizing constant. PIS learns a stochastic process that transports an easy initial distribution toward this target, turning sampling into a problem of controlling diffusion dynamics.
Learning a controlled diffusion
The construction starts from a Schrödinger bridge: a likely evolution between prescribed distributions. With a simple reference diffusion, the objective becomes a stochastic control problem with a running penalty on control effort and a terminal cost determined by the target density. A neural network represents the control and is trained end to end.
Sampling quality and correction
The analysis relates imperfect control to sampling error in Wasserstein distance. Path-integral importance weights provide a way to compensate for bias from an imperfect controller and time discretization. Together, these ideas connect learned sampling dynamics with both an optimization objective and a correction mechanism, without requiring the target’s normalization constant.
Publication
Path Integral Sampler: A Stochastic Control Approach for Sampling
Qinsheng Zhang and Yongxin Chen
ICLR 2022