Learning with Importance Weighted Variational Inference

09.05.2025 11:15 – 12:15

RESEARCH INSTITUTE FOR STATISTICS AND INFORMATION SCIENCE: STATISTICS SEMINAR


Several popular variational bounds involvingimportance weighting ideas have been proposed to generalize and improve on theEvidence Lower Bound (ELBO) in the context of maximum likelihood optimization,such as the Importance Weighted Auto-Encoder (IWAE) and the Variational Rényi(VR) bounds. The methodology to learn the parameters of interest using thesebounds typically amounts to running gradient-based variational inferencealgorithms that incorporate the reparameterization trick. However, the way thechoice of the variational bound impacts the outcome of variational inferencealgorithms can be unclear.

In this talk, we will present and motivate the VR-IWAEbound, a novel variational bound that unifies the ELBO, IWAE and VR boundsmethodologies. In particular, we will provide asymptotic analyses for theVR-IWAE bound and its reparameterized gradient estimator, which enable us tocompare of the ELBO, IWAE and VR bounds methodologies. Our work advances theunderstanding of importance weighted variational inference methods and we willillustrate our theoretical findings empirically.

Lieu

Bâtiment: Uni Mail

Boulevard du Pont-d'Arve 40
1205 Geneva

Room M 4220, 4th floor

entrée libre