Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-labelling via simultaneous clustering and representation 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 |
|---|---|---|---|
| add self-labelling iterations rather than training from initial random labels only (ablation) | SeLa on ImageNet with AlexNet · ImageNet linear probe, conv4 | top-1 accuracy: 18.3 -> 44.7 | Helped |
| increase self-labelling optimizations from 80 to 160 (ablation) | SeLa on ImageNet with AlexNet · ImageNet linear probe, conv4 | top-1 accuracy: 44.7 -> 44.6 | Hurt |
| increase the cluster count from 1k to 3k (ablation) | SeLa on ImageNet with AlexNet · ImageNet linear probe, conv4 | top-1 accuracy: 42.1 -> 44.7 | Helped |
| increase the cluster count beyond 3k (ablation) | SeLa on ImageNet with AlexNet · ImageNet linear probe, conv4 | top-1 accuracy: 44.7 -> 43.9 | Hurt |
| use 10 clustering heads instead of one (ablation) | SeLa on ImageNet with AlexNet · ImageNet linear probe, conv4 | Top-1: 44.7 -> 46.7 | Helped |
| use 10 clustering heads instead of one (ablation) | SeLa on ImageNet with ResNet-50 · ImageNet linear evaluation | Top-1: 51.8 -> 61.5 | Helped |
| replace Sinkhorn-Knopp label assignment with K-means (ablation) | SeLa on CIFAR-10 training data · CIFAR-10 full training data, CIFAR-10 kNN | accuracy: 72.9 -> 64.7 | Hurt |
| replace Sinkhorn-Knopp label assignment with K-means (ablation) | SeLa on CIFAR-10 training data · CIFAR-10 full training data, CIFAR-100 kNN | accuracy: 28.9 -> 19.3 | Hurt |
| train with 10 heads rather than one (ablation) | SeLa on ImageNet with ResNet-50 · ImageNet linear evaluation | Top-1: 51.8 -> 61.5 | Helped |
| use transferred ResNet-50 labels to train AlexNet rather than applying SeLa directly to AlexNet (ablation) | SeLa label transfer · ImageNet linear probe, AlexNet conv5 | top-1 accuracy: 39.6 -> 42.3 | Helped |
| add stronger augmentation during label-transfer retraining (ablation) | SeLa transferred labels on AlexNet · ImageNet linear probe, AlexNet conv5 | top-1 accuracy: 42.3 -> 44.7 | Helped |
| add stronger augmentation during label-transfer retraining (ablation) | SeLa transferred labels on AlexNet · Places linear probe, AlexNet conv4 | top-1 accuracy: 39.0 -> 37.9 | Hurt |
Closest papers by meaning
- Semi-Supervised Learning Made Simple With Self-Supervised Clustering (CVPR 2023 · similarity 0.89)
- Unsupervised Visual Representation Learning via Mutual Information Regularized Assignment (NeurIPS 2022 · similarity 0.89)
- Unsupervised Representation Transfer for Small Networks: I Believe I Can Distill On-the-Fly (NeurIPS 2021 · similarity 0.87)
- Beyond Learning Features: Training a Fully-Functional Classifier with ZERO Instance-Level Labels (AAAI 2022 · similarity 0.87)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.86)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (ICLR 2024 · similarity 0.85)
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.