Abstract

Concept-based models are an emerging paradigm in deep learning that constrains the inference process to operate through human-understandable variables, facilitating interpretability and human interaction. However, these architectures, on par with popular opaque neural models, fail to account for the true causal mechanisms underlying the target phenomena represented in the data. This hampers their ability to support causal reasoning tasks, limits out-of-distribution generalization, and hinders the implementation of fairness constraints. To overcome these issues, we present Causally reliable Concept Bottleneck Models (C2BMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. Finally, in light of the limitations of such approach, we anticipate our intended future directions while also discussing their relevance within the broader landscape of concept-based interpretability.

Arianna Casanova, Francesco De Santis, and Giovanni De Felice are academic researchers with a shared focus on concept-based interpretability. Arianna Casanova Flores is a postdoctoral researcher at the University of Liechtenstein. Previously, she received a PhD in Informatics at the Dalle Molle Institute for Artificial Intelligence (CH). Francesco De Santis is a PhD candidate at Politecnico di Torino (IT). He received a Master’s degree in Data Science from Politecnico di Torino. Giovanni De Felice is a postdoctoral researcher at Università della Svizzera italiana (CH). Previously, he completed his Ph.D. (2025) in the Data Mining & Machine Learning group at the University of Liverpool (UK). He is a co-creator of the open-source machine learning library PyTorch Concepts.