Post-hoc Uncertainty Learning Using a Dirichlet Meta-Model

It is known that neural networks have the problem of being over-confident when directly using output label distribution to generate uncertainty measures. Existing methods mainly resolve this issue by retraining entire model impose quantification capability so learned can achieve desired performance in accuracy and prediction simultaneously. However, training from scratch computationally expensive, a trade-off might exist between quantification. To end, we consider more practical post-hoc learning setting, where well-trained base given, focus on task at second stage training. We propose novel Bayesian approach Dirichlet meta-model, which effective efficient. Our proposed method requires no additional data flexible enough quantify different uncertainties easily adapt application settings, including out-of-domain detection, misclassification trustworthy transfer learning. Finally, demonstrate our meta-model approach's flexibility superior empirical these applications over multiple representative image classification benchmarks.