Fixed-Mean Gaussian Processes for Post-hoc Bayesian Deep Learning
Researchers have developed a new method called fixed-mean Gaussian Processes (FMGP) for estimating uncertainty in pre-trained deep neural networks. This approach fixes the Gaussian Process posterior mean to the DNN's output, allowing it to efficiently fit predictive variances without compromising accuracy. FMGP is architecture-agnostic and scales well to large datasets like ImageNet, offering improved uncertainty estimation and computational efficiency over existing methods. AI
IMPACT Provides a novel technique for improving the reliability of deep learning models by quantifying prediction uncertainty.