精準識別「界門綱目科屬種」！北大彭宇新團隊用細粒度樹先驗提升泛化

🔑 增強重點摘要

• •The research addresses the 'long-tail' distribution problem in biological datasets, where rare species lack sufficient training samples, by leveraging the structural constraints of the Linnaean taxonomy.
• •The method utilizes a hierarchical contrastive learning framework that forces the model to learn discriminative features at multiple granularities simultaneously, rather than just the leaf-node species level.
• •Experimental results demonstrate significant performance gains on standard benchmarks like iNaturalist, particularly in reducing 'top-down' classification errors where a model correctly identifies a species but misclassifies its higher-level family or order.

🛠️ 技術深入

• •Architecture: Incorporates a Tree-Structured Prior Module (TSPM) that injects taxonomic knowledge into the latent space of vision transformers (ViTs).
• •Loss Function: Employs a hierarchical cross-entropy loss combined with a tree-aware contrastive loss to enforce consistency across the taxonomic hierarchy.
• •Data Augmentation: Uses taxonomy-guided data augmentation strategies to synthesize features for underrepresented classes based on their phylogenetic proximity to well-represented classes.
• •Inference: Implements a top-down constrained decoding mechanism that prunes the search space during inference to ensure predicted labels adhere to valid biological paths.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線