A Framework For Explainable Deep Neural Models Using External Knowledge Graphs

Deep neural networks (DNNs) have become the gold standard for solving challenging classification problems, especially given complex sensor inputs (e.g., images and video). While DNNs are powerful, they are also brittle, and their inner workings are not fully understood by humans, leading to their use as \black-box" models. DNNs often generalize poorly when provided new data sampled from slightly shifted distributions; DNNs are easily manipulated by adversarial examples; and the decision-making process of DNNs can be difficult for humans to interpret. To address these challenges, we propose integrating DNNs with external sources of semantic knowledge. Large quantities of meaningful, formalized knowledge are available in knowledge graphs and other databases, many of which are publicly obtainable. But at present, these sources are inaccessible to deep neural methods, which can only exploit patterns in the signals they are given to classify. In this work, we conduct experiments on the ADE20K dataset, using scene classification as an example task where combining DNNs with external knowledge graphs can result in more robust and explainable models. We align the atomic concepts present in ADE20K (i.e., objects) to WordNet, a hierarchically-organized lexical database. Using this knowledge graph, we expand the concept categories which can be identified in ADE20K and relate these concepts in a hierarchical manner. The neural architecture we present performs scene classification using these concepts, illuminating a path toward DNNs which can efficiently exploit high-level knowledge in place of excessive quantities of direct sensory input. We hypothesize and experimentally validate that incorporating background knowledge via an external knowledge graph into a deep learning-based model should improve the explainability and robustness of the model.