On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation (original) (raw)
Fig 24
Examples of images with an increasing amount of flipped pixels and the corresponding predictions of the classifier.
Here pixels are flipped towards the class label given in parentheses above the heatmap. Pixels were flipped in steps of 1% of all pixels until the predicted class label changed. The plots show the output of the softmax function y and the output score of the preceding linear layer yp. The pixels were sorted before flipping in increasing order of the pixel-wise score, i.e. lowest scoring pixels were flipped first. In this panel the heatmap was computed for a classifier which did not produce the highest score, i.e. for a random false class label. The originally predicted label is given on the leftmost image in parentheses, the predicted label after the switch of the prediction is given in the rightmost image.