Research map / Label-efficient and robust learning / Semi-supervised learning
Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active 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 a feature re-alignment module (similarity loss plus strong-weak contrastive loss) that plugs into pseudo-label SSL frameworks so that any AL selection strategy works better in semi-supervised active learning.
Problem
Combining active learning and semi-supervised learning is hard: diversity- and uncertainty-based AL criteria do not transfer well to SSL with very few labels per class and strong-weak augmentations, and prior SSAL methods are tied to specific SSL algorithms.
Approach
Decomposes SSL risk into pseudo error and margin error, mapping them to diversity- and uncertainty-based AL, then adds a feature re-alignment module: a queue-based similarity loss clusters unlabeled weak-view features toward nearest labeled strong-view features, and a contrastive loss aligns weak and strong views so weak-view uncertainty approximates strong-view uncertainty.
Key results
- Wait, the above had an error. Correcting the call below.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature alignment (FA) to FixMatch | FixMatch + ViT · CIFAR100 #100 | accuracy (%): 45.27 -> 46.48 | Helped |
| Add feature alignment (FA) to FreeMatch | FreeMatch + ViT · CIFAR100 #100 | accuracy (%): 53.44 -> 60.93 | Helped |
| Add feature alignment (FA) to FixMatch | FixMatch + ViT · CIFAR100 #200 | accuracy (%): 64.86 -> 66.81 | Helped |
| Add feature alignment (FA) to FreeMatch | FreeMatch + ViT · CIFAR100 #200 | accuracy (%): 72.96 -> 76.17 | Helped |
| Add feature alignment (FA) to FixMatch | FixMatch + ViT · TinyImgnet #200 | accuracy (%): 38.55 -> 40.08 | Helped |
| Add feature alignment (FA) to FreeMatch | FreeMatch + ViT · TinyImgnet #200 | accuracy (%): 38.75 -> 40.73 | Helped |
| Add feature alignment (FA) to FixMatch | FixMatch + ViT · TinyImgnet #400 | accuracy (%): 51.14 -> 51.86 | Helped |
| Add feature alignment (FA) to FreeMatch | FreeMatch + ViT · TinyImgnet #400 | accuracy (%): 52.06 -> 52.79 | Helped |
| Add similarity loss Ls (ablation) | FreeMatch + ViT · CIFAR100 #100, CoreSet selection | accuracy (%): 56.43 -> 58.72 | Helped |
| Add similarity loss Ls (ablation) | FreeMatch + ViT · CIFAR100 #100, AlphaMix selection | accuracy (%): 51.43 -> 61.02 | Helped |
| Add similarity loss Ls (ablation) | FreeMatch + ViT · CIFAR100 #100, ActiveFT selection | accuracy (%): 53.65 -> 63.28 | Helped |
| Add similarity loss Ls (ablation) | FreeMatch + ViT · CIFAR100 #100, Noise selection | accuracy (%): 52.28 -> 56.81 | Helped |
Surprising findings
- Plain AL selection inside SSL often performs worse than random sampling.
- Performance keeps improving with larger strong-view queue length T up to about 10, with diminishing returns after.
Limitations
- Only one active-learning selection round is evaluated.
- Experiments use only two image datasets and two SSL algorithms.
- Diversity-based clustering is weak when only one or two labels per class exist, which the module only partly mitigates.
Experimental setup
Datasets: CIFAR-100 TinyImageNet
Benchmarks: CIFAR-100 TinyImageNet
Metrics: accuracy
Baselines: Random CoreSet AlphaMix ActiveFT Noise Supervised
Builds on: FixMatch FreeMatch CoreSet AlphaMix ActiveFT
Closest papers by meaning
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.89)
- Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.89)
- Instance-wise Supervision-level Optimization in Active Learning (CVPR 2025 · similarity 0.88)
- InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · similarity 0.87)
- Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios (NeurIPS 2024 · similarity 0.87)
- A Unified Framework for Heterogeneous Semi-supervised Learning (CVPR 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.