Research map / Label-efficient and robust learning / Self-supervised and contrastive
Barlow Twins: Self-Supervised Learning via Redundancy Reduction
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 |
|---|---|---|---|
| remove the redundancy-reduction term (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 57.3 | Hurt |
| remove the invariance term (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 0.1 | Hurt |
| normalize embeddings along the feature dimension (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 69.8 | Hurt |
| remove batch normalization from projector hidden layers (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 71.2 | No clear effect |
| remove projector hidden-layer batch normalization and embedding normalization (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 53.4 | Hurt |
| replace the Barlow Twins loss with cross-entropy with temperature (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 63.3 | Hurt |
| add stop-gradient to one branch (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 70.5 | Hurt |
| add a predictor network to one branch (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 70.2 | Hurt |
| add stop-gradient and a predictor network (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%): 71.4 -> 61.3 | Hurt |
| reduce batch size to 256 (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%) | No clear effect |
| increase projector output dimensionality (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%) | Helped |
| increase projector depth up to three layers (ablation) | Barlow Twins · ImageNet linear evaluation | Top-1 accuracy (%) | Helped |
Closest papers by meaning
- On the Stepwise Nature of Self-Supervised Learning (ICML 2023 · similarity 0.86)
- Bridging the Gap from Asymmetry Tricks to Decorrelation Principles in Non-contrastive Self-supervised Learning (NeurIPS 2022 · similarity 0.86)
- Exploring the Equivalence of Siamese Self-Supervised Learning via a Unified Gradient Framework (CVPR 2022 · similarity 0.86)
- Collapse-Proof Non-Contrastive Self-Supervised Learning (ICML 2025 · similarity 0.85)
- Information Flow in Self-Supervised Learning (ICML 2024 · similarity 0.84)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.83)
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.