Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Visual Representation Learning by Online Constrained K-Means
Check 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 |
|---|---|---|---|
| Relax balanced clustering to a minimum cluster-size ratio γ'=0.4 (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 63.1 -> 64.5 | Helped |
| Remove the minimum cluster-size constraint (γ'=0) (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 64.5 -> 41.3 | Hurt |
| Use last-epoch labels and centers instead of current-epoch values (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 51.2 -> 64.5 | Helped |
| Use last-epoch labels rather than current-epoch labels (ablation) | CoKe with last-epoch centers · ImageNet, 200-epoch single-view linear classification | Acc%: 0.4 -> 64.5 | Helped |
| Increase the number of clusters from 1,000 to 3,000 (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 63.4 -> 64.5 | Helped |
| Train multiple clustering tasks with different cluster counts (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 64.5 -> 65.3 | Helped |
| Use batch size 256 rather than 1,024 (ablation) | CoKe · ImageNet, 200-epoch single-view linear classification | Acc%: 64.5 -> 64.2 | No clear effect |
| Begin moving-average label and center smoothing after epoch 160 rather than use single-stage training (ablation) | CoKe single-view · ImageNet, 200-epoch linear classification | Acc%: 65.3 -> 65.8 | Helped |
| Begin moving average earlier, after epoch 120 (ablation) | CoKe single-view · ImageNet, 200-epoch linear classification | Acc%: 65.8 -> 64.3 | Hurt |
| Set the two-view soft-label weight α to 0.2 rather than 0.1 (ablation) | CoKe two-view · ImageNet linear classification | Acc%: 73.9 -> 74.9 | Helped |
| Add a prediction MLP and combine its classification loss with the projection loss (ablation) | CoKe two-view · ImageNet linear classification | Acc%: 74.4 -> 74.9 | Helped |
| Optimize only the prediction-MLP loss, removing the projection-MLP loss (ablation) | CoKe two-view · ImageNet linear classification | Acc%: 74.9 -> 73.3 | Hurt |
Closest papers by meaning
- Jigsaw Clustering for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.89)
- ClusterFit: Improving Generalization of Visual Representations (CVPR 2020 · similarity 0.89)
- Clustering-friendly Representation Learning via Instance Discrimination and Feature Decorrelation (ICLR 2021 · similarity 0.88)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (ICLR 2024 · similarity 0.86)
- Neural Clustering based Visual Representation Learning (CVPR 2024 · similarity 0.86)
- Hierarchical Correlation Clustering and Tree Preserving Embedding (CVPR 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.