Compressed Active Subspaces for Scalable Bayesian Inference

Compressed Active Subspaces (CAS) enables Bayesian inference for high-dimensional models by constructing active subspaces within a parameter space reduced via structured isometric embeddings. This method avoids the memory overhead of storing full-dimensional gradients, allowing scalability to larger neural networks while preserving predictive uncertainty quantification.