Research map / Label-efficient and robust learning / Semi-supervised learning
SoftMatch: Addressing the Quantity-Quality Tradeoff in Semi-supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use truncated Gaussian confidence weighting instead of linear, quadratic, or untruncated weighting (ablation) | SoftMatch · CIFAR-10, 40 labels | top-1 error rate (%): 7.73 -> 4.91 | Helped |
| Use truncated Gaussian confidence weighting instead of linear weighting (ablation) | SoftMatch · SVHN-10, 40 labels | top-1 error rate (%): 15.27 -> 2.33 | Helped |
| Use truncated Gaussian confidence weighting instead of quadratic weighting (ablation) | SoftMatch · SVHN-10, 40 labels | top-1 error rate (%): 84.11 -> 2.33 | Helped |
| Replace truncated Gaussian weighting with truncated Laplacian weighting (ablation) | SoftMatch · CIFAR-10, 40 labels | top-1 error rate (%): 4.91 -> 5.30 | Hurt |
| Replace truncated Gaussian weighting with truncated Laplacian weighting (ablation) | SoftMatch · CIFAR-10, 250 labels | top-1 error rate (%): 4.82 -> 5.14 | No clear effect |
| Replace truncated Gaussian weighting with truncated Laplacian weighting (ablation) | SoftMatch · SVHN-10, 40 labels | top-1 error rate (%): 2.33 -> 3.12 | Hurt |
| Replace adaptive global Gaussian parameters with fixed μ and σ² (ablation) | SoftMatch · CIFAR-10, 40 and 250 labels; SVHN, 40 labels | error rate | Hurt |
| Replace global Gaussian parameters with per-class parameter estimation (ablation) | SoftMatch · CIFAR-10, 40 labels | error rate | Hurt |
| Add Uniform Alignment to the Gaussian weighting method (ablation) | SoftMatch · CIFAR-10, 250 labels | error rate | Helped |
| Use per-class estimation instead of global Gaussian estimation (ablation) | SoftMatch · SVHN | error rate | Hurt |
| Use an empirically estimated uniform target distribution for alignment instead of a uniform target (ablation) | SoftMatch with Uniform Alignment · Long-tailed image classification | error rate | Hurt |
| Use SoftMatch rather than FlexMatch | Semi-supervised image classification · CIFAR-100, 400 labels | top-1 error rate (%): 39.94 -> 37.10 | Helped |
Closest papers by meaning
- Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images (AAAI 2022 · similarity 0.92)
- FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020 · similarity 0.90)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.89)
- Bi-Level Optimization for Semi-Supervised Learning with Pseudo-Labeling (AAAI 2025 · similarity 0.88)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.88)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.88)
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.