ETD

• explainable artificial intelligence
• few-shot learning
• interpretable machine learning
• metric-learning
• siamese networks

Few-Shot Learning (FSL) enables models to recognize new categories from only a handful of examples by learning transferable similarity functions. Despite its success, the decision mechanisms of metric-based FSL models, such as Siamese and Relation Networks, remain opaque: they can tell whether two samples are similar, but not why. This lack of interpretability limits their reliability, diagnostic value, and trustworthiness in real-world applications. This thesis investigates how interpretability can be achieved in metric-based FSL, addressing three research questions: (i) Can traditional explainability methods faithfully describe how few-shot metric models make decisions? (ii) How can post-hoc explainers be redesigned to capture pairwise and relational reasoning? (iii) Can interpretability be embedded directly into model architectures so that explanations are produced by design? To answer these questions, this thesis develops both post-hoc explainability and interpretability by-design approaches. For post-hoc explainability, two dedicated explainers are proposed: SINEX, which attributes similarity in Siamese Networks to image regions that support or contradict a match, and RENEX, which extends this idea to Relation Networks by explaining how query–support interactions shape relational scores. In experiments, traditional explainers such as Grad-CAM and LRP were applied as-is, while LIME and Integrated Gradients were adapted to operate on the full metric architecture. The proposed methods outperform all these baselines, providing more faithful and human-aligned explanations of pairwise reasoning. For interpretability-by-design, three complementary architectures are introduced. SonicProtoPNet extends the "this looks like that" part-based prototype reasoning to the audio domain, Pairwise Distance Trees replace opaque learned distances with transparent, rule-based regressors, and SiameseConceptNetworks embeds interpretability directly into the metric space by representing each image through human-defined semantic concepts. This last model enables explanations in natural language (i.e., "same bill shape, different head pattern") and maintains competitive few-shot performance across benchmarks. This dissertation advances three key points. First, post-hoc explainability techniques can be effective when specifically designed for metric-learning models in FSL, often overcoming limitations of standard explainability tools that require tailored adaptations. Second, moving toward by-design solutions that incorporate human-like reasoning objects (i.e., prototypical parts, concepts, decision-path rules) is advantageous for producing explanations that are useful to end users. Third, combining pairwise strategies with concept-based representations provides a practical path to making few-shot Siamese Networks transparent and robust. Overall, this thesis provides new tools and principles for building few-shot models whose relational reasoning is transparent, human-aligned, and suitable for trustworthy real-world deployment.