Weighted Mutual Information for Out-Of-Distribution Detection

Out-of-distribution detection has become an important theme in machine learning (ML) field, since the recognition of unseen data either "similar" or not (in- or out-of-distribution) to the ones the ML system has been trained on may lead to potentially fatal consequences. Operational data compliance with the training data has to be verified by the data analyst, who must also understand, in operation, if the autonomous decision-making is still safe or not. In this paper, we study an out-of-distribution (OoD) detection approach based on a rule-based eXplainable Artificial Intelligence (XAI) model. Specifically, the method relies on an innovative metric, i.e., the weighted mutual information, able to capture the different way decision rules are used in case of in- and OoD data.