Research map / Label-efficient and robust learning / Recognition and long-tail
What Do You See in Common? Learning Hierarchical Prototypes over Tree-of-Life to Discover Evolutionary Traits
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HComP-Net, a hierarchical prototypical network over the tree of life that learns evolutionary-trait prototypes at every internal node, using over-specificity, discriminative, and masking modules to avoid over-specific prototypes.
Problem
Discovering evolutionary traits from images as a hierarchy of prototypes over a phylogenetic tree, where existing prototype methods learn over-specific prototypes at internal nodes, fail to exclude features shared with contrasting clades, and cannot abstain from common prototypes at high tree levels
Approach
Extends ProtoPNet/PIP-Net-style prototype learning to a phylogenetic hierarchy with an over-specificity loss (prototypes must activate for every descendant species), a discriminative loss against contrasting clades, and a learned Gumbel-softmax masking module that flags over-specific prototypes without pruning them during training.
Key results
- HComP-Net reaches 70.01% fine-grained accuracy on Bird vs 36.18% for HPnet, and 97.35% on Butterfly, outperforming baselines there.
- Part purity on Bird rises from 0.14 (HPnet) to 0.77 with over-specificity loss and masking (HComP-Net full).
- On four unseen Bird species, HComP-Net outperforms HPnet at one level above leaf (e.g., Indigo Bunting 96.67% vs 49.2%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HComP-Net instead of the strongest listed baseline | Bird fine-grained classification · Bird | fine-grained accuracy: 84.23 -> 70.01 | Hurt |
| use HComP-Net instead of the strongest listed baseline | Butterfly fine-grained classification · Butterfly | fine-grained accuracy: 97.87 -> 97.35 | Hurt |
| use HComP-Net instead of the strongest listed baseline | Fish fine-grained classification · Fish | fine-grained accuracy: 93.58 -> 90.80 | Hurt |
| use HComP-Net instead of the strongest listed baseline | Spider fine-grained classification · Spider | fine-grained accuracy: 84.05 -> 76.19 | Hurt |
| use HComP-Net instead of the strongest listed baseline | Turtle fine-grained classification · Turtle | fine-grained accuracy: 69.56 -> 58.26 | Hurt |
| use HComP-Net instead of HPnet | classification of unseen Bird species · Fish Crow | fine-grained accuracy: 10.55 -> 53.33 | Helped |
| use HComP-Net instead of HPnet | classification of unseen Bird species · Rock Wren | fine-grained accuracy: 10.22 -> 53.33 | Helped |
| use HComP-Net instead of HPnet | classification of unseen Bird species · Indigo Bunting | fine-grained accuracy: 49.2 -> 96.67 | Helped |
| use HComP-Net instead of HPnet | classification of unseen Bird species · Bohemian Waxwing | fine-grained accuracy: 44.9 -> 70.00 | Helped |
| add the over-specificity loss (ablation) | HComP-Net on Bird · Bird | fine-grained accuracy: 58.32 -> 70.01 | Helped |
| add the discriminative loss (ablation) | HComP-Net on Bird · Bird | fine-grained accuracy: 65.99 -> 70.01 | Helped |
| add self-supervised contrastive loss (ablation) | HComP-Net on Bird · Bird | part purity: 0.53 $\pm$ 0.18 -> 0.72 $\pm$ 0.19 | Helped |
Surprising findings
- HPnet's accuracy did not improve with higher-resolution (28x28) feature maps, dropping to 20.68% on Bird.
- On Bird, HComP-Net's accuracy (70.01%) trails plain ConvNeXt-tiny (84.23%) and PIP-Net (83.38%), since the goal prioritizes interpretable common traits over accuracy.
Limitations
- Accuracy trails strong non-hierarchical baselines such as ConvNeXt-tiny and PIP-Net on Bird, Spider and Turtle.
- Trait descriptions for prototypes are based on human interpretation, and the method targets local traits only, excluding global contours.
- Results on Spider and Turtle are weak (76.19% and 58.26%), and the mask module requires tuning of lambda_ovsp that trades off accuracy.
Experimental setup
Datasets: CUB-200-2011 Fish dataset (38 species, PhyloNN preprocessing) Heliconius butterfly collection (Cambridge Butterfly) iNat2021 Araneae subset iNat2021 Testudines subset
Benchmarks: CUB-200-2011 (Bird subset) Fish Butterfly Spider Turtle
Metrics: fine-grained leaf accuracy (path probability) part purity accuracy on unseen species % masked prototypes
Baselines: ResNet-50 ConvNeXt-tiny ProtoPNet INTR PIP-Net HPnet
Builds on: ProtoPNet PIP-Net HPnet ProtoTree
Closest papers by meaning
- Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery (NeurIPS 2025 · similarity 0.85)
- BioCLIP: A Vision Foundation Model for the Tree of Life (CVPR 2024 · similarity 0.85)
- BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning (NeurIPS 2025 · similarity 0.84)
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes (NeurIPS 2026 · similarity 0.84)
- PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders (ICLR 2025 · similarity 0.84)
- PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.