Explaining deep convolutional models by measuring the influence of interpretable features in image classification

Abstract The accuracy and flexibility of Deep Convolutional Neural Networks (DCNNs) have been highly validated over the past years. However, their intrinsic opaqueness is still affecting reliability limiting application in critical production systems, where black-box behavior difficult to be accepted. This work proposes EBAnO , an innovative explanation framework able analyze decision-making process DCNNs image classification by providing prediction-local class-based model-wise explanations through unsupervised mining knowledge contained multiple convolutional layers. provides detailed visual numerical thanks two specific indexes that measure features’ influence precision process. has experimentally evaluated, both quantitatively qualitatively, (i) analyzing its with four state-of-the-art DCNN architectures, (ii) comparing results three strategies (iii) assessing effectiveness easiness understanding human judgment, means online survey. released as open-source code it freely available online.