A1777
Title: Batch and match: Black-box variational inference with a score-based divergence
Authors: Diana Cai - Flatiron Institute (United States) [presenting]
Abstract: Variational inference (VI) has become a popular approach for posterior inference in latent-variable models, and recent advances in black-box VI (BBVI) methods -enabled by advances in automatic differentiation- have made VI widely accessible in practice. However, standard BBVI often converges slowly for expressive variational families due to high-variance stochastic gradients. To address this limitation, the Batch-and-Match (BaM) algorithm performs variational inference by matching the scores of the variational and target distributions over batches of samples. BaM admits closed-form updates for full-covariance Gaussians and requires significantly fewer gradient evaluations than standard BBVI. Extensions to high-dimensional settings and richer variational families are also discussed.
