Research map / Label-efficient and robust learning / Recognition and long-tail
Label Hierarchy Transition: Delving Into Class Hierarchies to Enhance Deep Classifiers
Renzhen Wang, De Cai, Kaiwen Xiao, Xixi Jia, Xiao Han, Deyu Meng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical cross-entropy loss with per-level branches (ablation) | ResNet-50 · CUB-200-2011 Species (R=90%) | ACC: 55.9 -> 58.3 | Improved |
| add transition networks (ablation) | ResNet-50 + HCL · CUB-200-2011 Species (R=90%) | ACC: 58.3 -> 66.7 | Improved |
| add confusion loss (ablation) | ResNet-50 + HCL + Transition Networks · CUB-200-2011 Species (R=90%) | ACC: 66.7 -> 68.8 | Improved |
| learn transition matrices via transition networks instead of trainable parameters (ablation) | LHT · CUB-200-2011 Species (R=90%) | ACC: 58.1 -> 68.8 | Improved |
| vary hyper-parameter s of transition network (ablation) | LHT · CUB-200-2011 | ACC | No clear effect |
| increase confusion loss weight lambda from 0 to 0.01 (ablation) | LHT · CUB-200-2011 | ACC | Improved |
| use a much larger confusion loss weight lambda (ablation) | LHT · CUB-200-2011 semi-supervised | ACC | Hurt |
| apply LHT to ConvNeXt-Base | ConvNeXt-Base baseline · CUB-200-2011 Species (R=90%) | ACC: 73.3 -> 80.1 | Improved |
| apply LHT to Swin-Base | Swin-Base baseline · CUB-200-2011 Species (R=90%) | ACC: 72.4 -> 78.9 | Improved |
| apply LHT when two finest levels lack labels | HD-CNN baseline · CUB-200-2011 Species (R^2=90%) | ACC: 51.0 -> 60.4 | Improved |
| apply LHT with fully supervised labels | HRN · CUB-200-2011 Species (R=0%) | ACC: 86.7 -> 87.0 | Improved |
| apply LHT at heavy label missing | best baseline · Stanford Cars Maker (R=90%) | ACC: 59.7 -> 71.4 | Improved |
Closest papers by meaning
- Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Hierarchical Multi-Label Classification Networks (ICML 2018 · similarity 0.87)
- Hier-COS: Making Deep Features Hierarchy-aware via Composition of Orthogonal Subspaces (CVPR 2026 · similarity 0.86)
- Use All the Labels: A Hierarchical Multi-Label Contrastive Learning Framework (CVPR 2022 · similarity 0.86)
- Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery (NeurIPS 2025 · similarity 0.85)
- HiGDA: Hierarchical Graph of Nodes to Learn Local-to-Global Topology for Semi-Supervised Domain Adaptation (AAAI 2025 · similarity 0.84)
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.
