19 Oct 2024 · Santiago de Compostela, Spain
TELL was one of the earlier steps in this line of work, and it still captures the idea very clearly. The model uses non-negative weights and a thresholding mechanism so that the learned computation can be converted into logic rules. That means the network is not only trained to classify, but also to produce an explanation that can be read as a compact symbolic description.
The interesting part for me is that TELL is not limited to toy settings. It can be applied to real-valued inputs through an automatic thresholding head, which makes it relevant for concept bottlenecks and prototype-based models as well. In practice, that broadens the kinds of tasks where a logic-based explanation can be useful.
Why this matters for explainability
Many explanation methods are local, fragile, or tied to a single prediction. TELL was designed to provide more global structure. That makes it easier to discuss how the model behaves overall, and it creates a more stable bridge between the neural model and the rules a human can inspect.
What we observed
Across the experiments, the logic-based formulation stayed competitive while offering a cleaner interpretation pathway. That balance between performance and transparency is what made the result interesting to present at ECAI, and it remains one of the core ideas behind later work in this research line.
View the TELL code and check the AIHub article for a more accessible summary.