Variational Inference for DPGMM with Coresets

Performing estimation and inference on massive datasets under time and memory constraints is a critical task in machine learning. One approach to tackle these challenges is offered by coresets, succinct data summaries that come with strong theoretical guarantees, and can operate under computational resource restrictions. In this work, we explore how such data summaries can be used in posterior inference through variational methods. We develop a novel coreset construction for approximate posterior inference in the nonparametric Dirichlet process Gaussian mixture model. We empirically demonstrate how our method allows trading small approximation error for large gains in runtime and memory usage.