Fixed-Mean Gaussian Processes for Post-hoc Bayesian Deep Learning

ArXi:2412.04177v2 Announce Type: replace Recently, there has been an increasing interest in performing post-hoc uncertainty estimation about the predictions of pre-trained deep neural networks (DNNs). Given a pre-trained DNN via back-propagation, these methods enhance the original network by adding output confidence measures, such as error bars, without compromising its initial accuracy. In this context, we