Bayesian Pseudocoreset

Bayesian pseudocoresets are compact synthetic datasets designed to approximate the posterior distribution of a large dataset, enabling efficient Bayesian inference, especially for complex models like deep neural networks. Current research focuses on developing methods to construct these pseudocoresets by minimizing various divergence measures between the full data posterior and the posterior based on the synthetic dataset, often working within a function space to mitigate challenges posed by high dimensionality and multi-modality. This approach offers significant potential for reducing computational costs in Bayesian inference, particularly in applications like federated learning, while maintaining accurate posterior approximations and well-calibrated uncertainty estimates.