Failure Prediction by Confidence Estimation of Uncertainty-Aware Dirichlet Networks

Reliably assessing model confidence in deep learning and predicting errors\nlikely to be made are key elements in providing safety for model deployment, in\nparticular for applications with dire consequences. In this paper, it is first\nshown that uncertainty-aware deep Dirichlet neural networks provide an improved\nseparation between the confidence of correct and incorrect predictions in the\ntrue class probability (TCP) metric. Second, as the true class is unknown at\ntest time, a new criterion is proposed for learning the true class probability\nby matching prediction confidence scores while taking imbalance and TCP\nconstraints into account for correct predictions and failures. Experimental\nresults show our method improves upon the maximum class probability (MCP)\nbaseline and predicted TCP for standard networks on several image\nclassification tasks with various network architectures.\n