Declared failure over silent failure: abstention, explain ability and an abstention-aware evaluation framework

Abstract

Deep image classifiers are constructed in a way so that every input produces a label, whatever the model has learned about that input. The resulting failures are silent, e.g., a classifier confronted with an Out-Of-Distribution (OOD) or adversarially perturbed image returns a class assignment with no signal that the assignment should not be trusted. Selective classification yields a partial remedy by permitting the model to decline but its standard evaluation practice, e.g., computing performance over the accepted subset and reporting coverage separately, charges nothing for that decline. Therefore, a reported score improves monotonically as the model abstains more. This study makes two contributions. The first is empirical where a SelectiveNet architecture is compared against a matched baseline sharing the same convolutional backbone across four datasets (MNIST, CIFAR-10, chest X-ray for pneumonia detection and Cats vs. Dogs), evaluated, under Fast Gradient Sign Method (FGSM) perturbation and with Gradient-weighted Class Activation Mapping (Grad-CAM) applied to the selection head, as well as, the prediction head so that the abstention decision is itself explained. The second is the Abstention-Aware Evaluation Framework (AAEF) which extends the confusion matrix with an abstention column preserving every sample and defines three metrics on it. Boundedness, exact reduction to the classical metrics at zero abstention and correct extreme-case behavior are established and verified. Absolute error counts fell on all four datasets with missed pneumonia cases dropping from 17 to 7, though 4 to 14 deferrals were required per error avoided. Adversarial results were consistent, as well. As perturbation strength rises the baseline CNN degrades steadily, while SelectiveNet’s selective accuracy recovers because it lowers coverage and rejects the samples it cannot resolve. The optimism introduced by discarding abstentions ranged from 0.069 to 0.26 in balanced accuracy. Charging abstention as an outcome rather than an exclusion changes the reported performance of a selective classifier which is large enough to reverse a comparison that conventional evaluation would decide the other way.