Research map / Label-efficient and robust learning / Semi-supervised learning
Universum Prescription: Regularization Using Unlabeled Data
Xiang Zhang, Yann LeCun
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 dustbin-class universum prescription | 21-layer ConvNet · CIFAR-10, CIFAR-100 fine/coarse, STL-10 Tiny, ImageNet top-1/top-5 | test error (%): 7.02 -> 6.66; 37.58 -> 32.84; 22.74 -> 20.45; 31.16 -> 27.96; 34.39 -> 33.67; 13.68 -> 13.35 | Improved |
| add dustbin-class universum prescription | 21-layer ConvNet · STL-10 | test error (%): 31.16 -> 36.58 | Hurt |
| add uniform prescription | 21-layer ConvNet · CIFAR-100 fine/coarse and STL-10 Tiny | test error (%): 37.58 -> 36.23; 31.16 -> 30.15 | Improved |
| add uniform prescription | 21-layer ConvNet · CIFAR-10, STL-10, ImageNet top-1/top-5 | test error (%): 7.02 -> 7.59; 31.16 -> 36.54; 34.39 -> 34.61; 13.68 -> 13.70 | Hurt |
| use a fixed-threshold background class instead of the baseline | 21-layer ConvNet · CIFAR-10, CIFAR-100 fine/coarse, STL-10, ImageNet top-1/top-5 | test error (%): 7.02 -> 8.38; 37.58 -> 40.57; 22.74 -> 24.97; 31.16 -> 38.95; 34.39 -> 34.69; 13.68 -> 13.84 | Hurt |
| increase the probability of sampling unlabeled data, up to a point (ablation) | universum-prescription ConvNet · CIFAR-10, CIFAR-100, STL-10 | generalization gap | Improved |
| increase the unlabeled-data sampling probability too far (ablation) | universum-prescription ConvNet · CIFAR-10, CIFAR-100, STL-10 | test error | Hurt |
| use unlabeled data that may overlap supervised classes rather than universum data (ablation) | universum prescription on STL-10 · STL-10 | test error: 31.16 -> 36.58 | Hurt |
| use 80 million tiny images instead of STL-10's accompanying unlabeled data (ablation) | universum prescription on STL-10 Tiny · STL-10 Tiny | test error: 31.16 -> 27.96 | Improved |
| add dropout alongside universum prescription (ablation) | universum-prescription ConvNet · CIFAR-10, CIFAR-100, STL-10 | test error | No clear effect |
Closest papers by meaning
- Towards Unsupervised Domain Generalization (CVPR 2022 · similarity 0.86)
- Unbiased Subclass Regularization for Semi-Supervised Semantic Segmentation (CVPR 2022 · similarity 0.86)
- Unsupervised Deep Learning by Neighbourhood Discovery (ICML 2019 · similarity 0.86)
- UniCon: Combating Label Noise Through Uniform Selection and Contrastive Learning (CVPR 2022 · similarity 0.86)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.85)
- Out-Of-Domain Unlabeled Data Improves Generalization (ICLR 2024 · similarity 0.84)
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.
