Heatmap Assisted Accuracy Score Evaluation Method for Machine-Centric Explainable Deep Neural Networks

There have existed many studies about the explainable artificial intelligence (XAI) that explains logic behind complex deep neural network called a black box. At same time, researchers tried to evaluate explainability performance of various XAIs. However, most previous evaluation methods are human-centric, is, subjective, where they rely on how much results explanation similar what people’s decision is based rather than features actually affect in model. Their XAI selections also dependent datasets. Furthermore, focusing only output variation target class. On other hand, this paper proposes robust heatmap assisted accuracy score (HAAS) scheme over datasets helps selecting machine-centric algorithms show leads given classification network. The proposed method modifies input image with scores obtained by algorithm and then puts resultant (HA) images into estimate change. metric (HAAS) computed as ratio accuracies HA original images. verified models LeNet-5 for MNIST VGG-16 CIFAR-10, STL-10, ILSVRC2012 totally 11 saliency map, deconvolution, 9 layer-wise relevance propagation (LRP) configurations. Consequently, LRP1 LRP3, MINST showed largest HAAS values 1.0088 1.0079, CIFAR-10 achieved 1.1160 1.1254, STL-10 had 1.0906 1.0918, got 1.3207 1.3469. While consists ϵ-rules input, convolutional, fully-connected layers, LRP3 adopts bounded-rule an layer layers LRP1. consistency HAAS AOPC has been compared means Kullback-Leibler divergence, ensuring more independently since lower average divergence 0.0251 0.3048. In addition, validity further investigated through inverted test employs made up estimates degradation caused applying them experience biggest test.