Research map / Label-efficient and robust learning / Self-supervised and contrastive
ClusterFit: Improving Generalization of Visual Representations
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 |
|---|---|---|---|
| Apply ClusterFit to weakly supervised image pre-training | ResNet-50 trained on IG-ImageNet-1B · ImageNet-9K | top-1 accuracy: +4.6 points | Helped |
| Apply ClusterFit to weakly supervised image pre-training | ResNet-50 trained on IG-ImageNet-1B · iNaturalist | top-1 accuracy: +5.8 points | Helped |
| Apply ClusterFit to weakly supervised image pre-training | ResNet-50 trained on IG-ImageNet-1B · ImageNet-1K | top-1 accuracy: 1.5 points | Hurt |
| Apply ClusterFit to weakly supervised video pre-training | R(2+1)D-34 trained on IG-Verb-19M · Kinetics, Sports1M, and Sth-Sth V1 | top-1 accuracy: 3.2 - 4.3% | Helped |
| Apply ClusterFit to Jigsaw or RotNet | ResNet-50 self-supervised pre-training · ImageNet-1K, VOC07, Places205, and iNaturalist | transfer score: 3-5 points on Places205; 5-8 points on iNaturalist | Helped |
| Increase the cluster count K (ablation) | ClusterFit on Jigsaw and RotNet · ImageNet-1K, VOC07, Places205, and iNaturalist | transfer score | Helped |
| Train the self-supervised pre-training model for 2× longer instead of applying ClusterFit (ablation) | Jigsaw and RotNet ResNet-50 · ImageNet-1K, VOC07, Places205, and iNaturalist | transfer score | Mixed |
| Combine Jigsaw and RotNet cluster pseudo-label tasks (ablation) | ClusterFit ResNet-50 · ImageNet-1K | top-1 accuracy: more than 8 points | Helped |
| Use a higher-capacity pre-trained feature extractor for clustering (ablation) | R(2+1)D-18 ClusterFit student · Kinetics | transfer score: 2%-3% | Helped |
| Use unsupervised clustering instead of per-label clustering (ablation) | ClusterFit on IG-Verb-62M videos · Kinetics and Sports1M | transfer score | Helped |
| Increase pre-training label diversity from the top 10 verbs to all 438 verbs (ablation) | ClusterFit using IG-Verb video features · Kinetics | top-1 accuracy: around 9% | Helped |
| Apply ClusterFit iteratively (ablation) | ClusterFit · Transfer learning | transfer score | No clear effect |
Closest papers by meaning
- Joint Unsupervised Learning of Deep Representations and Image Clusters (CVPR 2016 · similarity 0.89)
- Unsupervised Visual Representation Learning by Online Constrained K-Means (CVPR 2022 · similarity 0.89)
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.89)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (ICLR 2024 · similarity 0.88)
- MOFI: Learning Image Representations from Noisy Entity Annotated Images (ICLR 2024 · similarity 0.87)
- Control-oriented Clustering of Visual Latent Representation (ICLR 2025 · similarity 0.86)
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.