On the Decrease of Grandmother-Entropy in AI

Maya Felix, Davi Geiger

Keywords: Entropy, ResNet, Confidence Measure

TL;DR: A novel information-theoretical model for deep neural networks with an application to confidence measure for classification.

Abstract: Machine intelligence is best achieved today through deep learning. We propose a method to quantify the information content of activation values in a neural network layer for a given input. First, we map the activation of each node to a probability that reflects the likelihood that this node is selected among all nodes in the layer, where the more activation, the higher the probability. Then, we measure the information of the layer across channels or features, using conditional entropy, and we refer to it as grandmother-entropy. Empirical evaluation of ResNet-50 shows that the grandmother-entropy decreases as the input image propagates forward through the network layers. Moreover, the grandmother-entropy of the last convolution layer provides a reliable confidence measure to the classification result.

Primary Area: interpretability and explainable AI

Submission Number: 10203