4Consorzio Interuniversitario Nazionale per l'Informatica, CINI, (Italy) 6Institut des Systèmes Complexes de Paris Île-de-France (ISC-PIF) - UAR 3611 CNRS, Paris (France) Deep learning methods are highly accurate, yet their opaque decision process prevents them from earning full human trust [1]. Concept-based models [2, 3] aim to address this issue by learning tasks based on a set of human-understandable concepts. However, state-of-the-art concept-based models rely on high-dimensional concept embedding representations which lack a clear semantic meaning, thus questioning the interpretability of their decision process [4, 5]. To overcome this limitation, we propose the Deep Concept Reasoner (DCR) [6], the first interpretable concept-based model that builds upon concept embeddings. In DCR framework, the utilization of neural networks does not involve direct task predictions. Instead, neural networks are employed to construct syntactic rule structures through the utilization of concept embeddings. These concept embeddings serve as a foundation for the subsequent rule evaluation, enabling the DCR to derive its final prediction. Notably, the evaluation of rules occurs based on the truth values associated with the concepts themselves rather than their embeddings. Consequently, the DCR framework upholds clear semantic principles, facilitating the provision of fully interpretable decision outcomes. The overarching process is illustrated in the accompanying Figure 1-left. One of the key advantages of DCR lies in its diferentiability, which enables its efective training as an independent module on concept databases.
Italy ∗Corresponding author. https://www.pietrobarbiero.eu/ (P. Barbiero) during training, and (iii), facilitates the generation of counterfactual examples providing the learnt rules as guidance.
AND
Concept Encoder
Concept-based Model
Generalization vs.
Interpretability yes
Interpretable CE+DCR (ours) CT+Decision Tree CT+Logistic Regression CE+XGBoost
Acknowledgments This work was supported by TAILOR and by HumanE-AI-Net, projects funded by EU Horizon 2020 research and innovation programme under GA No 952215 and No 952026, respectively.