Importance Weighted Variational Inference without the Reparameterization Trick

Kamélia Daudel, Minh-Ngoc Tran, Cheng Zhang · 2026

Read Importance Weighted Variational Inference without the Reparameterization Trick by Kamélia Daudel, Minh-Ngoc Tran, Cheng Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-01412.txt.

Description

Kamélia Daudel and co-authors dissect score-function estimators for importance-weighted variational inference, which approximates intractable densities via Monte Carlo objectives. Without reparameterization, optimization leans on REINFORCE estimators. Standard VIMCO estimators lose signal-to-noise as counts rise. The team builds a VIMCO mechanism dodging that collapse, including a High-Effective Sample Size estimator with square-root-of-N scaling. That estimator lacks zero variance at optimality, so the team adds VIMCO estimators reaching zero variance with tempering. Discrete latent-variable and state-space tests confirm the theory.

Catalog record

CREATORKamélia Daudel, Minh-Ngoc Tran, Cheng Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE173791
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01412

Browse the full library · About the Tower of Babel