Richer Bayesian Last Layers with Subsampled NTK Features
Researchers have developed a new method to improve Bayesian Last Layers (BLLs) for estimating uncertainty in neural networks. Their approach leverages a projection of Neural Tangent Kernel (NTK) features to account for variability across the entire network, addressing the underestimation of epistemic uncertainty found in standard BLLs. This method offers provably greater or equal posterior variances and includes a subsampling scheme to reduce computational costs. Empirical tests on various datasets showed improved calibration and uncertainty estimates compared to existing methods. AI
IMPACT Improves neural network calibration and uncertainty estimation, potentially leading to more reliable AI systems in critical applications.