Calibrated Sampling-Free Uncertainty Estimation in Bayesian Deep Learning

The article presents Calibrated Variance Propagation (CVP), a novel method for efficient uncertainty estimation in Bayesian deep learning, particularly addressing the challenges of modern architectures like transformers and CNNs. CVP enhances layer-wise analytical approximations of uncertainty, improving coverage metrics significantly—up to 14.6% for BEiT-3 on NLVR2 and 10.8% for ViLT on VQAv2—while maintaining computational efficiency comparable to Monte Carlo sampling. This advancement is crucial for practitioners as it enables reliable uncertainty quantification in deep learning applications without incurring the high costs associated with traditional sampling methods.