Research map / Label-efficient and robust learning / Self-supervised and contrastive
Towards Efficient and Effective Self-Supervised Learning of Visual Representations
Sravanti Addepalli, Kaushal Bhogale, Priyam Dey, R. Venkatesh Babu
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 |
|---|---|---|---|
| add rotation prediction auxiliary task | BYOL · CIFAR-10 (200 epochs) | linear accuracy: 89.30 -> 91.89 | Improved |
| add rotation prediction auxiliary task | BYOL · CIFAR-100 (200 epochs) | linear accuracy: 60.67 -> 67.03 | Improved |
| add rotation prediction auxiliary task | SimCLR · CIFAR-100 (200 epochs) | linear accuracy: 62.96 -> 65.40 | Improved |
| add rotation prediction auxiliary task | SwAV · CIFAR-10 (200 epochs) | linear accuracy: 83.60 -> 87.20 | Improved |
| add rotation prediction auxiliary task | SwAV (ResNet-50) · ImageNet-1k (30 epochs) | linear top-1 accuracy: 54.90 -> 57.30 | Improved |
| add both Jigsaw and rotation tasks (ablation) | BYOL · ImageNet-100 | linear top-1 accuracy: 71.02 -> 74.72 | Improved |
| use 12 rotation angles instead of 4 (ablation) | BYOL + rotation · CIFAR-100 | linear accuracy: 67.03 -> 63.61 | Hurt |
| use rotation set {0,180} instead of {90,270} (ablation) | BYOL + rotation · CIFAR-100 | linear accuracy: 62.96 -> 66.21 | Improved |
| share encoder and projection network with rotation task instead of only the first conv layer (ablation) | BYOL + rotation · CIFAR-100 | linear accuracy: 52.50 -> 67.03 | Improved |
| add rotation task when color jitter is removed (ablation) | BYOL · CIFAR-10 | linear accuracy: 85.90 -> 91.28 | Improved |
| use CE loss instead of SupCon loss for rotation task (ablation) | BYOL + rotation · CIFAR-10 | linear accuracy: 90.19 -> 91.89 | Improved |
| add rotation as an augmentation (invariance) instead of auxiliary task (ablation) | BYOL · CIFAR-10 | linear accuracy: 89.30 -> 79.70 | Hurt |
Closest papers by meaning
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.92)
- Self-Supervised Representation Learning by Rotation Feature Decoupling (CVPR 2019 · similarity 0.91)
- Scaling and Benchmarking Self-Supervised Visual Representation Learning (ICCV 2019 · similarity 0.91)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.90)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.89)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.88)
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.
