Research map / Label-efficient and robust learning / Recognition and long-tail
Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels
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 Twofold Debiasing (TFB), combining multi-layer feature fusion reconstruction, intermediate-layer feature alignment, and a prototype-based classifier debiasing with coarse-sample embeddings to learn fine-grained recognition from coarse labels in the C2FS task.
Problem
Coarse-to-Fine Few-Shot (C2FS) learning: training feature extractors with only coarse labels yields features lacking fine-grained detail (simplicity bias), and few-shot fine-grained classifiers overfit to biased distributions.
Approach
Enhance fine-grained detail in embeddings via multi-layer feature fusion reconstruction and alignment of intermediate-layer features (self-knowledge distillation), and calibrate fine-grained prototypes by augmenting support sets with nearest coarse-class training embeddings found via kNN-based coarse label assignment.
Key results
- On CIFAR-100, TFB reaches 84.15% 5-way 1-shot and 39.03% all-way 1-shot, versus 81.42% and 36.28% for PE-HCM.
- On LIVING-17, TFB reaches 91.89% 5-way and 53.42% all-way 1-shot, exceeding PE-HCM (90.94%, 53.09%).
- Ablation on LIVING-17 and CIFAR-100 shows the classifier debias module adds about 3-4 points in 5-way and about 4-5 points in all-way accuracy over the feature-only variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-layer feature fusion reconstruction (REC) (ablation) | TFB without REC, ALIGN, and CD · LIVING-17 and CIFAR-100 | accuracy: 89.45 -> 90.74; 45.59 -> 47.32; 77.67 -> 80.05; 33.83 -> 36.80 | Helped |
| Add intermediate layer feature alignment (ALIGN) after REC (ablation) | TFB with REC · LIVING-17 and CIFAR-100 | accuracy: 90.74 -> 91.06; 47.32 -> 48.73; 80.05 -> 81.79; 36.80 -> 37.41 | Helped |
| Add fine-grained classifier debiasing (CD) after REC and ALIGN (ablation) | TFB with REC and ALIGN · LIVING-17 and CIFAR-100 | accuracy: 91.06 -> 91.89; 48.73 -> 53.42; 81.79 -> 84.15; 37.41 -> 39.03 | Helped |
| Use training features without coarse-category restriction instead of kNN-selected coarse subsets (ablation) | TFB classifier debiasing · LIVING-17 and CIFAR-100 | accuracy: 91.89 -> 91.02; 53.42 -> 51.71; 84.15 -> 83.19; 39.03 -> 37.61 | Hurt |
| Use support-set coarse labels instead of kNN to identify coarse categories (ablation) | TFB classifier debiasing · LIVING-17 and CIFAR-100 | accuracy: 91.89 -> 92.91; 53.42 -> 54.95; 84.15 -> 86.92; 39.03 -> 43.22 | Helped |
Surprising findings
- Coarse-grained supervised pre-training yields models that outperform fine-grained supervised upper bounds in few-shot fine-grained settings, e.g., on CIFAR-100 the best prior method exceeds the fine upper-bound by 5.89% (5-way 1-shot).
- Intermediate layer representations retain more fine-grained information than the deepest layer under coarse supervision.
Limitations
- Performance is sensitive to the number n of added support samples; too many introduce false positives that distort the distribution.
- Relies on accurate coarse-label assignment via kNN over training features; coarse labels of support samples improve results substantially, suggesting dependence on this step.
Experimental setup
Datasets: BREEDS (LIVING-17, NONLIVING-26, ENTITY-13, ENTITY-30) CIFAR-100
Benchmarks: BREEDS (LIVING-17, NONLIVING-26, ENTITY-13, ENTITY-30) CIFAR-100
Metrics: accuracy (5-way 1-shot) accuracy (all-way 1-shot) 95% confidence interval over 1000 episodes
Baselines: MoCo-v2 MoCo-v2-ImageNet SWAV-ImageNet ANCOR SCGM-G SCGM-A PE-HCM Fine upper-bound
Builds on: MoCo-v2 ANCOR prototypical networks
Closest papers by meaning
- Fine-Grained Recognition: Accounting for Subtle Differences between Similar Classes (AAAI 2020 · similarity 0.88)
- Fine-Grained Angular Contrastive Learning With Coarse Labels (CVPR 2021 · similarity 0.87)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.86)
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.86)
- Cross-Layer and Cross-Sample Feature Optimization Network for Few-Shot Fine-Grained Image Classification (AAAI 2024 · similarity 0.86)
- Few-Shot Fine-Grained Image Classification with Progressively Feature Refinement and Continuous Relationship Modeling (AAAI 2025 · similarity 0.85)
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.