Black-box models, like deep neural networks, are unreliable in safety-critical applications, where predictions must be based on the correct input-level attribution and have to adhere to the intended human explanations. Recent advances in Explainable AI have led to studying Concept-based models (CBMs), architectures that integrate high-level concepts to guarantee “ante-hoc” explainability. However, existing CBMs suffer from several limitations that compromise their interpretability.In this talk, I will discuss the current problems of SotA CBMs and present our solution, GlanceNets, new CBMs that align the concepts to the intended semantics by leveraging causal representation learning and out-of-distribution recognition. I will also shape possible future directions for improving Concept-based models.
19/05/2023