Research map / Label-efficient and robust learning / Semi-supervised learning
InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised 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 InterLUDE, an SSL image classifier that fuses labeled and unlabeled embeddings via circular-shift embedding fusion and adds a cross-instance delta consistency loss, evaluated on CIFAR-10/100, STL-10 and the open-set Heart2Heart ultrasound benchmark.
Problem
Mainstream semi-supervised image classification losses combine a supervised loss with an unlabeled-only regularizer, leaving little direct interaction between labeled and unlabeled data; evaluation also often ignores open-set unlabeled data.
Approach
Interdigitated batch layout with circular-shift embedding fusion (each embedding is linearly blended with its neighbor, which is typically from the other data source), plus a loss that matches the average change in predictions under weak-vs-strong augmentation between labeled and unlabeled samples.
Key results
- On STL-10 with 40 labels and ViT-base, InterLUDE achieves 3.14% error versus 14.91% for the next best method (InstanT).
- On CIFAR-10 with 40 labels and CNN, InterLUDE reaches 4.51% error and InterLUDE+ 4.46%, reported as surpassing the fully-supervised result.
- On Heart2Heart, InterLUDE attains 13.75% error on CAMUS and 4.86% on Unity, the best among compared methods; FlatMatch underperforms (28.93% on CAMUS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add embedding fusion (ablation) | InterLUDE · CIFAR-10, 40 labels, CNN | error rate (%): 4.82$\pm$0.12 -> 4.51$\pm0.01$ | Helped |
| Add delta consistency loss (ablation) | InterLUDE · CIFAR-10, 40 labels, CNN | error rate (%): 5.00$\pm0.35$ -> 4.51$\pm0.01$ | Helped |
| Add embedding fusion (ablation) | InterLUDE · CIFAR-10, 40 labels, ViT | error rate (%): 2.31$\pm$0.9 -> 1.78$\pm$0.1 | Helped |
| Add delta consistency loss (ablation) | InterLUDE · CIFAR-10, 40 labels, ViT | error rate (%): 2.37$\pm0.86$ -> 1.78$\pm$0.1 | Helped |
| Add embedding fusion (ablation) | InterLUDE+ · CIFAR-10, 40 labels, ViT | error rate (%): 1.78$\pm$0.1 -> 1.55$\pm$0.1 | Helped |
| Add delta consistency loss (ablation) | InterLUDE+ · CIFAR-10, 40 labels, ViT | error rate (%): 1.57$\pm$0.1 -> 1.55$\pm$0.1 | No clear effect |
| Add SAT and SAF | InterLUDE · CIFAR-10, 40 labels, CNN | error rate (%): 4.51$\pm0.01$ -> 4.46$\pm$0.11 | Helped |
| Add SAT and SAF | InterLUDE · TMED-2 | error rate (%): 3.45$\pm$0.39 -> 3.25$\pm$0.17 | Helped |
| Use high labeled-unlabeled-interaction interdigitated layout instead of low-interaction layout (ablation) | InterLUDE embedding fusion · CIFAR-10, low-label regime | error rate | Helped |
| Increase delta consistency loss coefficient to very high values (ablation) | InterLUDE · Extremely low-label settings | performance | Hurt |
| Increase embedding fusion strength toward the 0.5 upper limit (ablation) | InterLUDE · CIFAR-10 | performance | Hurt |
Surprising findings
- FlatMatch, competitive on classic benchmarks, substantially underperforms on the open-set Heart2Heart medical benchmark.
- Interdigitated layouts with more labeled-unlabeled interactions help mainly in the low-label regime; the advantage diminishes with 4000 labels.
Limitations
- Delta consistency loss relies on average behavior across unlabeled instances; finer per-class designs are left for future work.
- Theoretical understanding of why embedding fusion helps is lacking.
- Open-set performance is not the design target; results on CAMUS remain high-variance for all methods.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 STL-10 TMED-2 Unity CAMUS
Benchmarks: CIFAR-10 CIFAR-100 STL-10 Heart2Heart (TMED-2, CAMUS, Unity)
Metrics: error rate (%) standard deviation over three runs
Baselines: FixMatch FlexMatch FreeMatch SoftMatch SimMatchV2 FlatMatch FlatMatch-e InstanT AdaMatch FeatMatch OpenMatch Fix-A-Step
Builds on: FixMatch Manifold MixUp SimMatch FlatMatch SAT/SAF (FlatMatch-related self-adaptive threshold and fairness)
Closest papers by meaning
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.90)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.90)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
- Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · similarity 0.87)
- Semi-Supervised Sparse Gaussian Classification: Provable Benefits of Unlabeled Data (NeurIPS 2024 · similarity 0.87)
- Revisiting Semi-Supervised Learning in the Era of Foundation Models (NeurIPS 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.