Скачать с ютуб ICML 2024: Differentiable Annealed Importance Sampling Minimizes The JS-Divergence (Zenn, Bamler) в хорошем качестве

ICML 2024: Differentiable Annealed Importance Sampling Minimizes The JS-Divergence (Zenn, Bamler)

Accepted paper at ICML 2024 by Johannes Zenn and Robert Bamler. PDF: https://openreview.net/pdf?id=rvaN2P1rvC Poster: https://icml.cc/media/PosterPDFs/ICML... We will present the paper at ICML in Vienna, Austria in Hall C 4-9, poster stand 2610, on Thursday 25 July, 11:30 am CEST. Hope to see you there! Abstract: Differentiable annealed importance sampling (DAIS), proposed by Geffner & Domke (2021) and Zhang et al. (2021), allows optimizing, among others, over the initial distribution of AIS. In this paper, we show that, in the limit of many transitions, DAIS minimizes the symmetrized KL divergence (Jensen-Shannon divergence) between the initial and target distribution. Thus, DAIS can be seen as a form of variational inference (VI) in that its initial distribution is a parametric fit to an intractable target distribution. We empirically evaluate the usefulness of the initial distribution as a variational distribution on synthetic and real-world data, observing that it often provides more accurate uncertainty estimates than standard VI (optimizing the reverse KL divergence), importance weighted VI, and Markovian score climbing (optimizing the forward KL divergence). Video Chapters: 0:00 Hello 0:23 Differentiable Annealed Importance Sampling 0:58 Theorem and Overview of Our Contributions 2:13 Empirical Results 1: Mass Covering Behavior 3:34 Empirical Results 2: Logistic Regression and GP Regression 4:30 Conclusions