Understanding Softmax Confidence and Uncertainty
This addresses uncertainty estimation in neural networks for AI safety and reliability, but is incremental as it builds on existing OOD detection research.
The paper investigates why softmax confidence sometimes correlates with epistemic uncertainty despite neural networks often failing to increase uncertainty on out-of-distribution data, identifying implicit biases like decision boundary structure and network filtering effects, and finds that overlap in final-layer representations is a key failure reason, reduced by pre-trained networks.
It is often remarked that neural networks fail to increase their uncertainty when predicting on data far from the training distribution. Yet naively using softmax confidence as a proxy for uncertainty achieves modest success in tasks exclusively testing for this, e.g., out-of-distribution (OOD) detection. This paper investigates this contradiction, identifying two implicit biases that do encourage softmax confidence to correlate with epistemic uncertainty: 1) Approximately optimal decision boundary structure, and 2) Filtering effects of deep networks. It describes why low-dimensional intuitions about softmax confidence are misleading. Diagnostic experiments quantify reasons softmax confidence can fail, finding that extrapolations are less to blame than overlap between training and OOD data in final-layer representations. Pre-trained/fine-tuned networks reduce this overlap.