Post-hoc Uncertainty Calibration for Domain Drift Scenarios
This work is significant for researchers and practitioners who rely on calibrated uncertainty estimates from deep learning models, especially in real-world applications where domain drift is common, by highlighting and mitigating a critical failure mode.
This paper addresses the problem of uncertainty calibration in deep neural networks, specifically under domain shift. It demonstrates that existing post-hoc calibration methods become over-confident under domain shift and proposes a simple perturbation strategy for validation sets that substantially improves calibration in such scenarios across various architectures and tasks.
We address the problem of uncertainty calibration. While standard deep neural networks typically yield uncalibrated predictions, calibrated confidence scores that are representative of the true likelihood of a prediction can be achieved using post-hoc calibration methods. However, to date the focus of these approaches has been on in-domain calibration. Our contribution is two-fold. First, we show that existing post-hoc calibration methods yield highly over-confident predictions under domain shift. Second, we introduce a simple strategy where perturbations are applied to samples in the validation set before performing the post-hoc calibration step. In extensive experiments, we demonstrate that this perturbation step results in substantially better calibration under domain shift on a wide range of architectures and modelling tasks.