SGD for Variational Inference: Tackling Unbounded Variance via Preconditioning and Dynamic Batching
arXiv:2605.07531v1 Announce Type: new
Abstract: Black-Box Variational Inference (BBVI) typically relies on Stochastic Gradient Descent (SGD) to optimize the Evidence Lower Bound (ELBO). However, the stochastic gradients in BBVI inherently exhibit unbo…