Differentiable Particle Filtering
Published in Oral (long talk), ICML 2021, 2020
A novel, principled approach to Differentiable Particle Filtering, using Optimal Transport. Long talk/ oral at ICML 2021.
Paper: https://arxiv.org/abs/2102.07850
Cite
∘∈proceed∈gs{corenflos2021d⇔erentiab≤,tit≤={D⇔erentiab≤partic≤fi<er∈gviaentropy-rega̲rizedoptimaltransport},author={Corenflos,AdrienandThorn→n,JamesandDeligia∩idis,Geor≥andDoucet,Arnaud},b∞ktit≤={InternationalConferenceonMaχ≠Learn∈g},pa≥s={2100--2111},year={2021},organization={PMLR}}Particle Filtering (PF) methods are an established class of procedures for performing inference in non-linear state-space models. Resampling is a key ingredient of PF, necessary to obtain low variance likelihood and states estimates. However, traditional resampling methods result in PF-based loss functions being non-differentiable with respect to model and PF parameters. In a variational inference context, resampling also yields high variance gradient estimates of the PF-based evidence lower bound. By leveraging optimal transport ideas, we introduce a principled differentiable particle filter and provide convergence results. We demonstrate this novel method on a variety of applications. Accepted as a Long talk at ICML 2021.
