Research map / Graph learning / Graph self-supervised learning
Unifying Visual Contrastive Learning for Object Recognition from a Graph Perspective
Shixiang Tang, Feng Zhu, Lei Bai, Rui Zhao, Chenyu Wang, Wanli Ouyang
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 |
|---|---|---|---|
| use non-linear ReLU(BN) activation instead of identity in the predictor (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 68.5 -> 69.3 | Improved |
| make W' trainable under ReLU(BN) activation (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 67.5 -> 69.3 | Improved |
| use Softmax instead of ReLU(BN) activation (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 69.3 -> 69.4 | No clear effect |
| add neighborhood aggregation on the target branch (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 69.3 -> 71.4 | Improved |
| add self-loop term on the target branch (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 71.4 -> 71.8 | Improved |
| add aggregation and self-loop terms to online branch as well (ablation) | GCN predictor · ImageNet-1K linear eval (200 epochs, predictor ablation) | top-1 accuracy: 71.8 -> 71.9 | No clear effect |
| add node feature masking with drop probability 0.3 (ablation) | UniVCL · ImageNet-1K linear eval (200 epochs, graph augmentation) | top-1 accuracy: 71.9 -> 72.1 | Improved |
| add edge modification with drop probability 0.3 (ablation) | UniVCL · ImageNet-1K linear eval (200 epochs, graph augmentation) | top-1 accuracy: 71.9 -> 72.2 | Improved |
| raise graph masking drop probability to 0.7 (ablation) | UniVCL · ImageNet-1K linear eval (200 epochs, graph augmentation) | top-1 accuracy: 71.9 -> 70.1 | Hurt |
| add PPR diffusion on the target branch (ablation) | UniVCL · ImageNet-1K linear eval (200 epochs, graph augmentation) with diffusion | top-1 accuracy: 72.2 -> 72.8 | Improved |
| add heat diffusion online and PPR diffusion target (ablation) | UniVCL · ImageNet-1K linear eval (200 epochs, graph augmentation) with diffusion | top-1 accuracy: 72.2 -> 72.9 | Improved |
| add neighborhood aggregation graph predictor vs MOCOv2 identity predictor (200 epochs) | UniVCL · ImageNet-1K linear eval (200 epochs) | top-1 accuracy: 68.6 -> 72.9 | Improved |
Closest papers by meaning
- MA-GCL: Model Augmentation Tricks for Graph Contrastive Learning (AAAI 2023 · similarity 0.89)
- Return to Basics: Very Simple Graph Contrastive Learning via Noise Cancellation Principle (NeurIPS 2026 · similarity 0.89)
- Graph Contrastive Learning with Augmentations (NeurIPS 2020 · similarity 0.88)
- Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning (NeurIPS 2023 · similarity 0.88)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.86)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · 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.
