Generalized Laplace Active Subspaces for Scalable Uncertainty Quantification

A new low-rank generalized Laplace approximation enables efficient uncertainty quantification in neural networks by focusing on a small number of data-informed curvature directions. This method allows for closed-form posterior variances in a subspace, bypassing the computational cost of full Bayesian inference.