| Abstract |
We present a distributed information fusion system able to integrate heterogeneous information processing services based on machine learning and reasoning approaches. We focus on higher (semantic) levels of information fusion, and highlight the requirement for the component services, and the system as a whole, to generate explanations of its outputs. Using a case study approach in the domain of traffic monitoring, we introduce component services based on (i) deep neural network approaches and (ii) heuristic-based reasoning. We examine methods for explanation generation in each case, including both transparency (e.g, saliency maps, reasoning traces) and post-hoc methods (e.g, explanation in terms of similar examples, identification of relevant semantic objects). We consider trade-offs in terms of the classification performance of the services and the kinds of available explanations, and show how service integration offers more robust performance and explainability. |