Stefano Gamba
Hierarchical Concept Representation in Concept Bottleneck Models via Box Embeddings.
Rel. Eliana Pastor. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) | Preview |
Abstract
As Deep Neural Networks are increasingly deployed in safety-critical domains, the need for Explainable Artificial Intelligence (XAI) continues to grow. Concept Bottleneck Models (CBMs) offer an interpretable-by-design architecture by mapping raw inputs to intermediate, human-readable semantic concepts before producing final predictions. However, standard CBM paradigms predominantly rely on flat linear classifiers, making them vulnerable to shortcut learning and concept leakage, while also overlooking hierarchical relationships among concepts. Sequential optimization can lead to algebraic compensation and reliance on spurious concept co-occurrences, while joint end-to-end training allows classification gradients to propagate through the bottleneck, potentially encoding task-specific visual information that is not aligned with the intended concept semantics.
To address these limitations, this thesis introduces a novel decoupled architecture that explicitly models hierarchical relationships among attributes using Box Embeddings
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
