Finding the input features that reduce the entropy of a neural network’s prediction

Abstract

In deep learning-based image classification, the entropy of a neural network’s output is often taken as a measure of its uncertainty. We introduce an explainability method that identifies those features in the input that impact most this uncertainty. Learning the corresponding features by straightforward backpropagation typically leads to results that are hard to interpret. We propose an extension of the recently proposed oriented, modified integrated gradients (OMIG) technique as an alternative to produce perturbations of the input that have a visual quality comparable to explainability methods from the literature but marks features that have a substantially higher impact on the entropy. The potential benefits of the modified OMIG method are demonstrated by comparison with current state-of-the-art explainability methods on several popular databases. In addition to a qualitative analysis of explainability results, we propose a metric for their quantitative comparison, which evaluates the impact of identified features on the entropy of a prediction.