Salient Explanation for Fine-Grained classification

Abstract

Explaining the prediction of deep models has gained increasing attention to increase its applicability, even spreading it to life-affecting decisions. However there has been no attempt to pinpoint only the most discriminative features contributing specifically to separating different classes in a fine-grained classification task. This paper introduces a novel notion of salient explanation and proposes a simple yet effective salient explanation method called Gaussian light and shadow (GLAS), which estimates the spatial impact of deep models by the feature perturbation inspired by light and shadow in nature. GLAS provides a useful coarse-to-fine control benefiting from scalability of Gaussian mask. We also devised the ability to identify multiple instances through recursive GLAS. We prove the effectiveness of GLAS for fine-grained classification using the fine-grained classification dataset. To show the general applicability, we also illustrate that GLAS has state-of-the-art performance at high speed (about 0.5 sec per 224\times 224 image) via the ImageNet Large Scale Visual Recognition Challenge.