Year: 2025
Type: conference-paper
Abstract: Identifying the most appropriate metric to capture similarities between embeddings remains a longstanding challenge in Machine Learning (ML).Recent research has highlighted the potential of non-Euclidean geometries for modeling embedding distances, particularly in datasets with latent hierarchical structures.However, selecting an optimal geometry is non-trivial, and the literature lacks a comprehensive analysis of the advantages and limitations associated with different metric spaces.In this paper, we aim to address this gap by focusing on a setting where the choice of geometry plays a crucial role in the optimization process: Prototype Learning (PL).Through extensive comparisons across a diverse range of datasets, we uncover key insights into the behavior of non-Euclidean spaces, showcasing their limitation and possible future developments.
Authors: Fonio, Samuele, Grosso, Silvia, Bouchenak, Sara
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7140017725
FWCI:
Cites:
Cited by:
Related to: