Skip to content

Differentiable Annealed Importance Sampling and the Perils of Gradient Noise #11

@Kashalpha

Description

@Kashalpha

一言でいうと

Annealed importance sampling (AIS)とHamiltonian Monte Carlo (HMC)を合体させた手法(Differentiable AIS, DAIS)を提案することで、AISによる周辺尤度の推定量が微分不可能となる問題を回避した。

論文リンク

https://proceedings.neurips.cc/paper/2021/hash/a1a609f1ac109d0be28d8ae112db1bbb-Abstract.html

概要

  • DAISではHMCのように運動量を補助変数として導入し、AISにおけるMetropolis-Hastings法による受容判断を行わずに、leapfrog法による決定論的な遷移と運動量のrefreshmentにより状態の変化を行う。この状態の変化に加えて、AISでの逆温度のアニーリングを行う。DAISによるLower boundは順過程と逆過程のKL divergenceとして得られる。
  • 推定量が微分可能となったことで、変分推論に活用することができる。

先行研究

コメント

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions