Amortizing Tempered Sampling with Corrected Self-Training
Sequential Score Sampler matches competing samplers’ accuracy while cutting runtime and energy evaluations on several tested targets.
The authors train a neural sampler across a sequence of tempered distributions, correcting its samples with importance sampling and Markov chain Monte Carlo. Replay buffers retain samples for further training. Their Sequential Score Sampler uses a score-based diffusion model trained on energy gradients, with importance weights correcting its proposed sampling paths. Once trained, it can supply samples directly, through corrected proposals, or from the final-temperature buffer.
ArXiv · AI/CL/LG's note
The authors train a neural sampler across a sequence of tempered distributions, correcting its samples with importance sampling and Markov chain Monte Carlo. Replay buffers retain samples for further training. Their Sequential Score Sampler uses a score-based diffusion model trained on energy gradients, with importance weights correcting its proposed sampling paths. Once trained, it can supply samples directly, through corrected proposals, or from the final-temperature buffer.
ArXiv · AI/CL/LG's note
score 4