Calibration of Model Uncertainty for Dropout Variational Inference

verfasst von

Max-Heinrich Laves, Sontje Ihler, Karl-Philipp Kortmann, Tobias Ortmaier

Abstract

The model uncertainty obtained by variational Bayesian inference with Monte Carlo dropout is prone to miscalibration. In this paper, different logit scaling methods are extended to dropout variational inference to recalibrate model uncertainty. Expected uncertainty calibration error (UCE) is presented as a metric to measure miscalibration. The effectiveness of recalibration is evaluated on CIFAR-10/100 and SVHN for recent CNN architectures. Experimental results show that logit scaling considerably reduce miscalibration by means of UCE. Well-calibrated uncertainty enables reliable rejection of uncertain predictions and robust detection of out-of-distribution data.

Organisationseinheit(en)

Institut für Mechatronische Systeme

Typ

Preprint

Publikationsdatum

20.06.2020

Publikationsstatus

Elektronisch veröffentlicht (E-Pub)

Elektronische Version(en)

https://arxiv.org/abs/2006.11584 (Zugang: Offen)