Towards Neural Network Interpretability Using Commonsense Knowledge Graphs

Youmna Ismaeil, Daria Stepanova, Trung-Kien Tran, Piyapat Saranrittichai, Csaba Domokos, Hendrik Blockeel

Abstract: Convolutional neural networks (CNNs) classify images by learning intermediate representations of the input throughout many layers. In recent work, latent representations of CNNs have been aligned with semantic concepts. However, for generating such alignments, the majority of existing methods predominantly rely on large amounts of labeled data, which is hard to acquire in practice. In this work, we address this limitation by presenting a framework for mapping hidden units from CNNs to semantic attributes of classes extracted from external commonsense knowledge repositories. We empirically demonstrate the effectiveness of our framework on copy-paste adversarial image classification and generalized zero-shot learning tasks.