Research map / Label-efficient and robust learning / Semi-supervised learning
Roll with the Punches: Expansion and Shrinkage of Soft Label Selection for Semi-supervised Fine-Grained Learning
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 SoC, a soft-label selection method for semi-supervised fine-grained visual classification that uses class-transition-tracking k-medoids clustering and confidence-aware k selection, reporting gains over FixMatch and other SSL baselines on Semi-Aves and Semi-Fungi.
Problem
Pseudo-labeling in semi-supervised fine-grained visual classification (SS-FGVC) is unreliable because fine-grained unlabeled data yields low-accuracy hard pseudo-labels that propagate errors.
Approach
Select a per-sample subset of classes for soft pseudo-labels by k-medoids clustering in class space, using a similarity derived from class prediction transitions across epochs (Class Transition Tracking), with the number of clusters k set by prediction confidence; shrinking the selected set is shown to reduce entropy.
Key results
- On Semi-Fungi trained from scratch, SoC reaches 39.4% Top-1 vs. 32.2% for FixMatch and 36.0% for FlexMatch.
- With MoCo pre-training, MoCo + SoC reaches 47.2% Top-1 on Semi-Fungi from scratch, versus 44.2% for MoCo + FlexMatch.
- On Semi-Aves from ImageNet initialization, SoC reaches 57.8% Top-1, slightly above FixMatch's 57.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SoC instead of the strongest reported baseline | Semi-Aves and Semi-Fungi SS-FGVC · Semi-Aves and Semi-Fungi, from scratch, ImageNet, and iNat | Top-1 accuracy: 31.3 vs 28.1; 57.8 vs 57.4; 71.3 vs 70.2; 39.4 vs 36.5; 61.4 vs 59.6; 62.4 vs 60.1 | Helped |
| Initialize SoC with MoCo pretraining | SoC · Semi-Aves and Semi-Fungi, from scratch, ImageNet, and iNat | Top-1 accuracy: 31.3 -> 39.3; 57.8 -> 58.0; 71.3 -> 70.8; 39.4 -> 47.2; 61.4 -> 61.9; 62.4 -> 62.5 | Mixed |
| Replace confidence-aware k selection with any tested fixed k (ablation) | SoC · Semi-Fungi, from scratch | Top-1 accuracy: 39.4 -> 36.5; 39.4 -> 35.8; 39.4 -> 36.0; 39.4 -> 36.6 | Hurt |
| Set α to K/(K-2), 2, or 10 in the linear confidence-to-k mapping (ablation) | SoC · Semi-Fungi, from scratch | Top-1 accuracy: 39.4 -> 36.6; 39.4 -> 37.8; 39.4 -> 35.8 | Hurt |
| Replace the linear confidence-to-k mapping with the tested exponential mapping (ablation) | SoC · Semi-Fungi, from scratch | Top-1 accuracy: 39.4 -> 34.9; 39.4 -> 36.7; 39.4 -> 37.2; 39.4 -> 36.9 | Hurt |
| Replace SoC soft-label selection with FixMatch at the tested confidence thresholds (ablation) | SoC · Semi-Fungi, from scratch | Top-1 accuracy: 39.4 vs 34.7; 39.4 vs 34.3; 39.4 vs 30.3; 39.4 vs 26.15 | Hurt |
| Replace CTT clustering with labeled-data class centroids (ablation) | SoC · Semi-Fungi with OOD unlabeled data, from scratch | Top-1 accuracy: 39.4 -> 34.5 | Hurt |
| Replace CTT clustering with pseudo-label class centroids (ablation) | SoC · Semi-Fungi with OOD unlabeled data, from scratch | Top-1 accuracy: 39.4 -> 35.5 | Hurt |
Surprising findings
- Soft labels with lower pseudo-label accuracy still outperform hard FixMatch-style pseudo-labels during training.
- Gains from SoC over FixMatch are small when training from ImageNet-pretrained models on Semi-Aves (57.8 vs 57.4 Top-1).
Limitations
- Gains over the strongest baselines are small when pre-trained models are used (e.g., ~0.7% Top-1 on Semi-Aves from ImageNet).
- Hyperparameters such as alpha and the CTT averaging window were tuned on the benchmark and the entropy-reduction proof relies on conditions that the authors only argue hold in practice.
Experimental setup
Datasets: Semi-Aves Semi-Fungi ImageNet iNaturalist 2018 CIFAR-10
Benchmarks: Semi-Aves Semi-Fungi
Metrics: Top-1 accuracy Top-5 accuracy
Baselines: Pseudo-Label Curriculum Pseudo-Label FixMatch FlexMatch KD-Self-Training SimMatch MoCo MoCo + FlexMatch
Builds on: FixMatch FlexMatch SimMatch MoCo k-medoids
Closest papers by meaning
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
- Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning (ICML 2024 · similarity 0.86)
- Bi-Level Optimization for Semi-Supervised Learning with Pseudo-Labeling (AAAI 2025 · similarity 0.86)
- Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning (AAAI 2024 · similarity 0.85)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.85)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · 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.