Research map / Label-efficient and robust learning / Recognition and long-tail
SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery
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 SEAL, a multi-level GCD framework that uses naturally occurring semantic hierarchies (e.g., genus, car type) with cross-granularity self-distillation and hierarchical semantic-guided soft contrastive learning to reach SOTA on fine-grained GCD benchmarks.
Problem
Generalized Category Discovery must classify unlabelled images from both known and novel classes; prior hierarchical GCD methods rely on abstract or manually designed hierarchies that may be noisy, and standard contrastive losses treat all negatives equally.
Approach
Use real semantic taxonomies (publicly available or LLM-generated) as multi-level supervision, align predictions across granularities via a learned dynamic transition matrix with KL self-distillation, and replace InfoNCE with a soft contrastive loss whose targets reflect hierarchical similarity and a hybrid cosine/distance metric.
Key results
- With DINOv2 on SSB, average All accuracy 76.3% vs 74.9% for DebGCD (previous best).
- On Herbarium19, All accuracy 46.9% vs 45.8% for μGCD (previous best).
- Ablation on Stanford Cars/CUB/Aircraft: full model improves over baseline by +11.5, +5.9, +8.6 All accuracy (per the paper's own table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hierarchical learning to the baseline (ablation) | GCD on Stanford Cars · Stanford Cars (SSB) | ACC All: 53.8 -> 57.5 | Helped |
| Add hierarchical learning to the baseline (ablation) | GCD on CUB · CUB (SSB) | ACC All: 60.3 -> 57.0 | Hurt |
| Add consistency self-distillation to hierarchical learning (ablation) | GCD on Stanford Cars · Stanford Cars (SSB) | ACC All: 57.5 -> 62.6 | Helped |
| Add consistency self-distillation to hierarchical learning (ablation) | GCD on CUB · CUB (SSB) | ACC All: 57.0 -> 57.8 | Helped |
| Add semantic-guided HSCL to hierarchical learning (ablation) | GCD on Stanford Cars · Stanford Cars (SSB) | ACC All: 57.5 -> 64.4 | Helped |
| Add semantic-guided HSCL to hierarchical learning (ablation) | GCD on CUB · CUB (SSB) | ACC All: 57.0 -> 62.5 | Helped |
| Combine hierarchical learning, consistency self-distillation, and semantic-guided HSCL (ablation) | GCD on Stanford Cars · Stanford Cars (SSB) | ACC All: 53.8 -> 65.3 | Helped |
| Combine hierarchical learning, consistency self-distillation, and semantic-guided HSCL (ablation) | GCD on CUB · CUB (SSB) | ACC All: 60.3 -> 66.2 | Helped |
| Replace semantic coarse-level labels with 100% random labels (ablation) | SEAL · CUB | ACC All: 66.2 -> 30.3 | Hurt |
| Replace semantic coarse-level labels with 100% random labels (ablation) | SEAL · Stanford Cars | ACC All: 65.3 -> 29.6 | Hurt |
| Replace semantic coarse-level labels with 100% random labels (ablation) | SEAL · FGVC-Aircraft | ACC All: 62.0 -> 33.2 | Hurt |
| Use τ_c=0.75 instead of τ_c=0.5 (ablation) | SEAL on Stanford Cars · Stanford Cars unlabelled set | ACC All: 62.9 -> 65.3 | Helped |
Surprising findings
- Randomly assigned coarse-level labels (100% or 50%) drop performance sharply on CUB, SCars and Aircraft, suggesting gains depend on semantic alignment of the hierarchy.
Limitations
- Evaluation limited to datasets with fewer than 700 classes, restricting category coverage.
- Underperforms DebGCD and InfoSieve on CUB, attributed to inconsistent human-annotated bird taxonomy.
- Gains on generic datasets are smaller because coarse groupings do not form a common hierarchy.
Experimental setup
Datasets: CUB Stanford Cars FGVC-Aircraft Oxford-Pet Herbarium19 CIFAR-10 CIFAR-100 ImageNet-100
Benchmarks: SSB (CUB, Stanford Cars, FGVC-Aircraft) Oxford-Pet Herbarium19 CIFAR-10 CIFAR-100 ImageNet-100
Metrics: ACC (All, Old, New)
Baselines: SimGCD GCD μGCD DebGCD InfoSieve CiPR SPTNet PromptCAL XCon OpenCon
Builds on: SimGCD GCD DINO DINOv2 InfoNCE
Closest papers by meaning
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.90)
- Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery (NeurIPS 2025 · similarity 0.89)
- Semantic-Guided Novel Category Discovery (AAAI 2024 · similarity 0.88)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.88)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · similarity 0.87)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.87)
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.