Research map / Label-efficient and robust learning / Self-supervised and contrastive
Efficient Unsupervised Visual Representation Learning with Explicit Cluster Balancing
Ioannis Maniadis Metaxas, Georgios Tzimiropoulos, Ioannis Patras
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 predictor head (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, multi-crop) | linear accuracy: 73.8 -> 74.0 | Improved |
| stop gradient to centroid layer for local crops (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, multi-crop) | linear accuracy: 74.0 -> 74.3 | Improved |
| cosine weight decay schedule 0.001 to 0.0001 for centroid layer (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, multi-crop) | linear accuracy: 74.3 -> 74.5 | Improved |
| reduce batch size from 1024 to 256 (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, no multi-crop) | linear accuracy: 70.7 -> 70.5 | No clear effect |
| vary number of clusters K away from 65536 (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, no multi-crop) | linear accuracy: 70.7 -> 70.3 (K=32768); 70.7 -> 70.6 (K=98304) | Hurt |
| use linear rather than quadratic balancing function for small clusters (ablation) | ExCB · ImageNet linear (ResNet50, 100 ep, multi-crop) | linear accuracy: 74.3 -> 74.5 | Improved |
| ExCB balancing vs DINO centering: fraction of empty clusters | ExCB · ImageNet train clusters (K=65,536) | empty clusters: 74.7% -> 1.6% | Improved |
| ExCB vs SMoG on semi-supervised 1% labels | ExCB · ImageNet semi-supervised 1% labels | top-1 accuracy: 58.0 -> 57.8 | No clear effect |
| ExCB vs MoCo-v2 on COCO detection | ExCB · COCO val2017 Mask R-CNN | AP^b: 39.3 -> 39.2 | No clear effect |
| ExCB vs prior methods with fewer epochs and smaller batch | ExCB · ImageNet linear (ResNet50) | linear top-1 accuracy: 76.4 -> 76.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MAST — MAST: A Memory-Augmented Self-Supervised Tracker (CVPR 2020) · 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
- Unsupervised Visual Representation Learning by Online Constrained K-Means (CVPR 2022 · similarity 0.88)
- Self-supervised Learning is More Robust to Dataset Imbalance (ICLR 2022 · similarity 0.87)
- Unsupervised Representation Learning by Balanced Self Attention Matching (ECCV 2024 · similarity 0.87)
- GOCA: Guided Online Cluster Assignment for Self-Supervised Video Representation Learning (ECCV 2022 · similarity 0.87)
- ClusterFit: Improving Generalization of Visual Representations (CVPR 2020 · similarity 0.87)
- Joint Unsupervised Learning of Deep Representations and Image Clusters (CVPR 2016 · 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.
