Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

Being Bayesian, Even Just a Bit, Fixes Overconfidence in ReLU Networks

About

The point estimates of ReLU classification networks---arguably the most widely used neural network architecture---have been shown to yield arbitrarily high confidence far away from the training data. This architecture, in conjunction with a maximum a posteriori estimation scheme, is thus not calibrated nor robust. Approximate Bayesian inference has been empirically demonstrated to improve predictive uncertainty in neural networks, although the theoretical analysis of such Bayesian approximations is limited. We theoretically analyze approximate Gaussian distributions on the weights of ReLU networks and show that they fix the overconfidence problem. Furthermore, we show that even a simplistic, thus cheap, Bayesian approximation, also fixes these issues. This indicates that a sufficient condition for a calibrated uncertainty on a ReLU network is "to be a bit Bayesian". These theoretical results validate the usage of last-layer Bayesian approximation and motivate a range of a fidelity-cost trade-off. We further validate these findings empirically via various standard experiments using common deep ReLU networks and Laplace approximations.

Agustinus Kristiadi, Matthias Hein, Philipp Hennig• 2020

Related benchmarks

TaskDatasetResultRank
Out-of-Distribution DetectionCIFAR-100
AUROC77.27
107
Out-of-Distribution DetectionCIFAR-10 vs SVHN (test)
AUROC0.9186
101
Out-of-Distribution DetectionCIFAR-100 SVHN in-distribution out-of-distribution (test)
AUROC80.09
90
Out-of-Distribution DetectionCIFAR-10 in-distribution LSUN out-of-distribution (test)
AUROC86.81
73
Out-of-Distribution DetectionCIFAR-100 (in-distribution) / LSUN (out-of-distribution) (test)
AUROC76.5
67
Out-of-Distribution DetectionSVHN
AUROC89.47
62
Out-of-Distribution DetectionSVHN CIFAR-10 in-distribution out-of-distribution (test)
AUROC80.73
56
Out-of-Distribution DetectionMNIST (In-distribution) vs Fashion-MNIST (OOD) (test)
AUPR0.9737
36
Out-of-Distribution DetectionLSUN
AUROC0.9077
26
Out-of-Distribution DetectionSVHN → CIFAR-100 (test)
AUROC88.77
22
Showing 10 of 18 rows

Other info

Follow for update