Steering Diffusion Models to Rare Events with Sequential Monte Carlo
DireSMC is pitched as a faster way to sample and price rare events from diffusion-model simulators.
The paper targets cases where ordinary Monte Carlo becomes impractical because the needed sample count grows as the event gets rarer. Its method steers weighted samples toward a user-defined rare event while producing a calibrated probability estimate. The authors report tests on a toy problem and a score-based climate emulator, estimating probabilities from `10^-3` to `10^-5`. They claim net speed-ups from `9x` to `1413x` over Monte Carlo. ArXiv · AI/CL/LG's note
The paper targets cases where ordinary Monte Carlo becomes impractical because the needed sample count grows as the event gets rarer. Its method steers weighted samples toward a user-defined rare event while producing a calibrated probability estimate. The authors report tests on a toy problem and a score-based climate emulator, estimating probabilities from `10^-3` to `10^-5`. They claim net speed-ups from `9x` to `1413x` over Monte Carlo. ArXiv · AI/CL/LG's note
score 5