Augmenting Softmax Information for Selective Classification with Out-of-Distribution Data

AbstractDetecting out-of-distribution (OOD) data is a task that receiving an increasing amount of research attention in the domain deep learning for computer vision. However, performance detection methods generally evaluated on isolation, rather than also considering potential downstream tasks tandem. In this work, we examine selective classification presence OOD (SCOD). That to say, motivation detecting samples reject them so their impact quality predictions reduced. We show under specification, existing post-hoc perform quite differently compared when only detection. This because it no longer issue conflate in-distribution (ID) with if ID going be misclassified. conflation within correct and incorrect becomes undesirable. propose novel method SCOD, Softmax Information Retaining Combination (SIRC), augments softmax-based confidence scores feature-agnostic information such ability identify improved without sacrificing separation between predictions. Experiments wide variety ImageNet-scale datasets convolutional neural network architectures SIRC able consistently match or outperform baseline whilst fail do so. Code available at https://github.com/Guoxoug/SIRC.