Details
| Original language | English |
|---|---|
| Number of pages | 8 |
| Publication status | E-pub ahead of print - 2019 |
Abstract
Keywords
- cs.LG, stat.ML
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
2019.
Research output: Working paper/Preprint › Preprint
}
TY - UNPB
T1 - Well-calibrated Model Uncertainty with Temperature Scaling for Dropout Variational Inference
AU - Laves, Max-Heinrich
AU - Ihler, Sontje
AU - Kortmann, Karl-Philipp
AU - Ortmaier, Tobias
N1 - Accepted at 4th workshop on Bayesian Deep Learning (NeurIPS 2019)
PY - 2019
Y1 - 2019
N2 - Model uncertainty obtained by variational Bayesian inference with Monte Carlo dropout is prone to miscalibration. The uncertainty does not represent the model error well. In this paper, temperature scaling is extended to dropout variational inference to calibrate model uncertainty. Expected uncertainty calibration error (UCE) is presented as a metric to measure miscalibration of uncertainty. The effectiveness of this approach is evaluated on CIFAR-10/100 for recent CNN architectures. Experimental results show, that temperature scaling considerably reduces miscalibration by means of UCE and enables robust rejection of uncertain predictions. The proposed approach can easily be derived from frequentist temperature scaling and yields well-calibrated model uncertainty. It is simple to implement and does not affect the model accuracy.
AB - Model uncertainty obtained by variational Bayesian inference with Monte Carlo dropout is prone to miscalibration. The uncertainty does not represent the model error well. In this paper, temperature scaling is extended to dropout variational inference to calibrate model uncertainty. Expected uncertainty calibration error (UCE) is presented as a metric to measure miscalibration of uncertainty. The effectiveness of this approach is evaluated on CIFAR-10/100 for recent CNN architectures. Experimental results show, that temperature scaling considerably reduces miscalibration by means of UCE and enables robust rejection of uncertain predictions. The proposed approach can easily be derived from frequentist temperature scaling and yields well-calibrated model uncertainty. It is simple to implement and does not affect the model accuracy.
KW - cs.LG
KW - stat.ML
U2 - 10.48550/arXiv.1909.13550
DO - 10.48550/arXiv.1909.13550
M3 - Preprint
BT - Well-calibrated Model Uncertainty with Temperature Scaling for Dropout Variational Inference
ER -