Research map / Label-efficient and robust learning / Semi-supervised learning
Image-Feature Weak-to-Strong Consistency: An Enhanced Paradigm for Semi-Supervised Learning
Zhiyu Wu, Jinshi Cui
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image-feature weak-to-strong consistency (IFMatch) to FixMatch | FixMatch · CIFAR-100 (400 labels) | accuracy (%): 53.58 -> 66.26 | Improved |
| apply IFMatch to FreeMatch on STL-10 with 40 labels | FreeMatch · STL-10 (40 labels) | accuracy (%): 84.44 -> 84.75 | Improved |
| apply IFMatch to FlexMatch on STL-10 1000 labels | FlexMatch · STL-10 (1000 labels) | accuracy (%): 94.23 -> 95.24 | Improved |
| apply IFMatch to imbalanced SSL | FixMatch · CIFAR-10-LT (gamma=150) | accuracy (%): 70.38 -> 75.59 | Improved |
| combine strong feature and strong image perturbation in one branch (ablation) | FixMatch · CIFAR-10 (40 labels) | accuracy (%): 92.53 -> 88.26 | Hurt |
| add confidence-based identification of naive samples for weak feature perturbation (ablation) | IFMatch (Fix) · CIFAR-100 (400 labels) | accuracy (%): 63.56 -> 66.26 | Improved |
| use confidence-based identification instead of SAA loss+OTSU (ablation) | IFMatch (Fix) · CIFAR-100 (400 labels) | accuracy (%): 65.32 -> 66.26 | Improved |
| use constant threshold 0.95 instead of dynamic threshold in feature-perturbation branch I (ablation) | IFMatch (Soft) · CIFAR-100 (400 labels) | accuracy (%): 63.80 -> 67.02 | Improved |
| include each feature-level perturbation strategy (dropout, translation, shearing, value) (ablation) | IFMatch (Fix) · CIFAR-100 (400 labels) | accuracy (%): 65.68 -> 66.26 (channel dropout) | Improved |
| apply IFMatch to semi-supervised segmentation | UniMatch · Cityscapes (1/16 labels) | mIoU: 75.0 -> 76.3 | Improved |
| add strong feature perturbation branch (training cost) (ablation) | IFMatch (Fix) · CIFAR-100-400 training speed | sec/iter: 0.2058 -> 0.2640 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- FixMatch — FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) · better on 6 of 6 shared results
- FreeMatch — FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023) · better on 2 of 2 shared results
- SoftMatch — SoftMatch: Addressing the Quantity-Quality Tradeoff in Semi-supervised Learning (ICLR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Flexible Distribution Alignment: Towards Long-tailed Semi-supervised Learning with Proper Calibration (ECCV 2024 · similarity 0.89)
- Revisiting Consistency Regularization for Semi-Supervised Learning (IJCV 2022 · similarity 0.89)
- ConMatch: Semi-Supervised Learning with Confidence-Guided Consistency Regularization (ECCV 2022 · similarity 0.89)
- Revisiting Weak-to-Strong Consistency in Semi-Supervised Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images (AAAI 2022 · similarity 0.89)
- FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020 · similarity 0.89)
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.
