Research map / Label-efficient and robust learning / Self-supervised and contrastive
Weak Augmentation Guided Relational Self-Supervised Learning
Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, Chang Xu
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 weak augmentation instead of contrastive augmentation for the teacher (ablation) | ReSSL · Tiny ImageNet | linear top-1: 38.66 -> 46.64 | Improved |
| use weak augmentation instead of contrastive augmentation for teacher (ablation) | ReSSL · CIFAR-100 | linear top-1: 58.60 -> 65.03 | Improved |
| add random resized crop to no teacher augmentation (ablation) | ReSSL · CIFAR-10 | linear top-1: 79.33 -> 90.62 | Improved |
| add color jitter, grayscale and blur on top of crop and flip for teacher (ablation) | ReSSL · CIFAR-100 | linear top-1: 65.03 -> 58.60 | Hurt |
| add InfoNCE warm-up (ablation) | ReSSL · Tiny ImageNet | linear top-1: 46.64 -> 48.22 | Improved |
| add InfoNCE warm-up (ablation) | ReSSL · CIFAR-10 | linear top-1: 90.76 -> 89.98 | Hurt |
| remove predictor (symmetric system) with tau_s=0.1 and tau_t>=0.08 (ablation) | ReSSL · CIFAR-10 | linear top-1: 90.20 best -> 10.00 collapsed | Hurt |
| add predictor head (ablation) | ReSSL · CIFAR-10 | linear top-1: 90.20 -> 90.76 | Improved |
| use tau_s=0.1 instead of 0.2 (ablation) | ReSSL · CIFAR-100 | linear top-1: 63.12 -> 65.03 | Improved |
| increase teacher temperature tau_t to 0.10 (flatter target) (ablation) | ReSSL · Tiny ImageNet | linear top-1: 46.64 -> 32.28 | Hurt |
| increase memory buffer from 256 to 16384 (ablation) | ReSSL · Tiny ImageNet | linear top-1: 42.80 -> 46.64 | Improved |
| increase memory buffer from 16384 to 32768 (ablation) | ReSSL · Tiny ImageNet | linear top-1: 46.64 -> 45.95 | Hurt |
Closest papers by meaning
- ReSSL: Relational Self-Supervised Learning with Weak Augmentation (NeurIPS 2021 · similarity 0.97)
- Directional Self-Supervised Learning for Heavy Image Augmentations (CVPR 2022 · similarity 0.89)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.87)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.86)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.86)
- Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation (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.
