The implementation of artificial intelligence models in high-stakes applications leads to a growing need for explainable artificial intelligence (XAI) in order to guarantee the validity of the inferences made by these systems.

AI models face an imposed trade-off between power and intelligibility. The most powerful models are black boxes that require the creation of an interface to explain their results. However, this interface, being external to the model, may not faithfully reflect its inner working. In addition, the explanations produced are often unsatisfactory for non-specialists. In contrast, intrinsic methods, based on models that are easier to understand, are generally less efficient in solving complex problems.

We propose to develop hybrid methods between opacity and transparency, based on the extraction of semantic features in order to benefit from both the power of black-box type models and the intelligibility of intrinsic methods. These methods would be applied to mine warfare, for the detection and classification of underwater mines from sonar imagery