papers/zoo-species-classification/sections/14-supplementary-prototype-visualization.tex

\section{Supplementary: Prototype Visualization}

\label{app:visualization}

Analysis of learned prototypes using t-SNE reveals clear taxonomic
clustering: species prototypes cluster tightly within genera, genera form
distinct sub-clusters within families, and families separate cleanly in
the embedding space. The taxonomic prototypical network naturally
organizes the embedding space to reflect biological relationships,
emerging from the hierarchical episode sampling and multi-rank loss.

Spatial attention maps show that TaxoNet focuses on taxonomically
informative features: wing bar patterns for warbler identification,
venation patterns for butterfly classification, leaf margin shapes for
plant identification, and operculum structure for marine gastropod
classification. These attention patterns are consistent with the
diagnostic characters used by human taxonomists.