Explanation in Machine Learning systems has been identi ed to be the main asset to have for large scale deployment of Arti cial Intelligence (AI) in critical systems. Explanations could be example-, features-, semantics-based or even counterfactual to potentially action on an AI system; they could be represented in many di erent ways e.g., textual, graphical, or visual. All representations serve di erent means, purpose and operators. We built the rst-of-its-kind XAI (eXplainable AI) platform for critical systems i.e., Thales XAI Platform which aims at serving explanations through various forms. This paper emphasizes on the semantics-based explanations for Machine Learning systems.
Critical Applications: From adapting a plane trajectory, stopping a train, re tting a boat to recon guring a satellite, all are examples of critical situations where explanation is a must-to-have to follow an AI system decision. 2
Why Knowledge Graphs for Explainable AI? State-of-the-Art Limitations: Most approaches limits explanation of ML systems to features involved in the data and model, or at best to examples, prototypes or counterfactuals. Explanation should go beyond correlation (features importance) and numerical similarity (local explanation).
Opportunity: By expanding and linking initial (training, validation and test) data with entities in knowledge graphs, (i) context is encoded, (ii) connections and relations are exposed, and (iii) inference and causation are natively supported. Knowledge graphs are used for encoding better representation of data, structuring a ML model in a more interpretable way, and adopt a semantic similarity for local (instance-based) and global (model-based) explanation. 3
Thales XAI Platform: A Knowledge Graph Perspective (Semantic) Perspective: The platform is combining ML and reasoning functionalities to expose a human-like rational as explanation when (i) recognizing an object (in a raw image) of any class in a knowledge graph, (ii) predicting a link in a knowledge graph. Thales XAI Platform is using state-of-the-art Semantic Web tools for enriching input, output (class) data with DBpedia (4; 233; 000 resources) and domain-speci c knowledge graphs, usually enterprise knowledge graphs. This is a crucial step for contextualizing training, validation, test data. Explainable ML Classi cations: Starting from raw images, as unstructured data, but with class labels augmented with a domain knowledge graph, Thales XAI Platform relies on existing neural network architectures to build the most appropriate models. All con dence scores of output classes on any input image are updated based on the semantic description of the output classes. For instance, an input classi ed as a car will have a higher overall con dence score in case some properties of car in the knowledge graph are retrieved e.g., having wheels, being on a road. In addition the platform is embedding naturally explanation i.e., properties of the objects retrieved in both the raw data and knowledge graph. Explainable Relational Learning: Starting from relational data, structured as graph, and augmented with a domain knowledge graph, Thales XAI Platform relies on existing knowledge graph embeddings frameworks to build the most appropriate models. Explanation of any link prediction is retrieved by identifying representative hotspots in the knowledge graph i.e., connected parts of the graphs that negatively impact prediction accuracy when removed.