Does Implicit Clustering Matter? Comparing Quality of Embedding-Space Explanations from Cross-Entropy and Triplet Loss Training

Belloni, Julia Eva (2025) Does Implicit Clustering Matter? Comparing Quality of Embedding-Space Explanations from Cross-Entropy and Triplet Loss Training. Bachelor's Thesis, Artificial Intelligence.

Text Finalthesis.pdf Restricted to Registered users only Download (7MB)

Abstract

Interpretable embedding spaces enhance model transparency, allowing practitioners to detect failure modes and spurious correlations early in development. In high-stakes applications, it is crucial to understand the model's reasoning to prevent incorrect decision-making. In the context of image classification, we compare embedding spaces learned with metric-based triplet loss to those produced by conventional cross-entropy training. Experiments were conducted on Imagenette using ResNet-18 backbones, with Grad-CAM, Eigen-CAM, and Guided Grad-CAM employed to evaluate explanation faithfulness and robustness. We found that triplet-loss models consistently produce more faithful explanations. However, both training strategies yield comparable predictive performance, computational cost, and explanation stability. These results highlight triplet loss as a strong alternative to cross-entropy, combining equivalent accuracy with superior interpretability.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Zullich, M. and Valdenegro Toro, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	15 Jul 2025 10:40
Last Modified:	15 Jul 2025 10:40
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36277

Actions (login required)

View Item