• model interpretability
• artificial neural networks
• machine learning
• artificial intelligence

Neural networks deployed in critical environments could benefit from the generation of sound explanations for their outputs.
As a first step in this direction, existing techniques can analyze the correlation between units in convolutional neural networks for image classification and the presence of visual concepts in their input.
By building on previous methods and formalizing visual concepts in the context of a referential theory of meaning, this work introduces a theoretical framework to estimate semantic alignment.
Furthermore, by considering visual concepts as members of an ontology, the proposed approach improves the quality of the alignment and enables the clustering of units into semantically coherent and architecturally connected circuits.
Circuits are experimentally evaluated in a classification context, highlighting how they select units critical for the accuracy of semantically related classes.