Hyperbolic Few-Shot Learning for Taxonomic Plant Classification

Mithil Shah

Submission Track: Short papers presenting ongoing research or work submitted to other venues (up to 5 pages, excluding references)

Keywords: few-shot learning, hyperbolic geometry, prototypical networks, taxonomy, biodiversity

TL;DR: HProtoNet uses hyperbolic geometry to improve few-shot plant classification by better capturing taxonomic hierarchies.

Abstract: Few-shot classification of biological species remains a challenging problem, especially when taxonomic hierarchies must be respected. In this paper, we investigate the role of hyperbolic geometry in modeling plant taxonomy within the PlantCLEF dataset. We introduce HProtoNet, a hyperbolic prototypical network variant, and compare it against Euclidean Prototypical Networks and Matching Networks. Through hierarchical accuracy analysis, few-shot comparisons, and Poincar\'e disk embedding visualizations, we demonstrate that hyperbolic embeddings better capture the inherent tree-like structure of species relationships. Our results highlight the promise of geometry-aware few-shot learning for biodiversity applications. We further argue that these methods not only improve classification but also align computational predictions with biological intuition, making them particularly suitable for ecological deployment.

Submission Number: 52