Resampling Gradients Vanish in Differentiable Sequential Monte Carlo Samplers

Johannes Zenn, Robert Bamler
4/27/2023
stat.MLcs.LG

No official code found. Best implementation: jzenn/DSMCS (5 stars, MIT, last commit 2024-05-21).

Abstract

Annealed Importance Sampling (AIS) moves particles along a Markov chain from a tractable initial distribution to an intractable target distribution. The recently proposed Differentiable AIS (DAIS) (Geffner and Domke, 2021; Zhang et al., 2021) enables efficient optimization of the transition kernels of AIS and of the distributions. However, we observe a low effective sample size in DAIS, indicating degenerate distributions. We thus propose to extend DAIS by a resampling step inspired by Sequential Monte Carlo. Surprisingly, we find empirically-and can explain theoretically-that it is not necessary to differentiate through the resampling step which avoids gradient variance issues observed in similar approaches for Particle Filters (Maddison et al., 2017; Naesseth et al., 2018; Le et al., 2018).

View on arXivView PDF

Code Implementations(1)

jzenn/DSMCSCommunity100%
MIT

Code of "Resampling Gradients Vanish in Differentiable Sequential Monte Carlo Samplers" (TinyPaper@ICLR'23)

51PythonApr 19, 20232 years agoMIT

Similar Papers

Differentiable Annealed Importance Sampling Minimizes The Symmetrized Kullback-Leibler Divergence Between Initial and Target Distribution

84% match1 repo42024

Diffusion differentiable resampling

84% match2025

Adaptive Annealed Importance Sampling with Constant Rate Progress

83% match2023

Variance Reduction of Resampling for Sequential Monte Carlo

82% match2023

Ensemble-Based Annealed Importance Sampling

81% match2024

Importance Weighted Variational Inference without the Reparameterization Trick

80% match2026

Cite this paper

@article{zenn2023resampling,
  title  = {Resampling Gradients Vanish in Differentiable Sequential Monte Carlo Samplers},
  author = {Johannes Zenn and Robert Bamler},
  year   = {2023},
  eprint = {2304.14390},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2304.14390v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2304.14390.svg)](https://codeofpaper.com/paper/2304.14390)

Discussion