Research map / Label-efficient and robust learning / Self-supervised and contrastive
SEED: Self-supervised Distillation For Visual Representation
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 |
|---|---|---|---|
| Distill the student from a larger self-supervised teacher rather than train it directly with MoCo-V2 | EfficientNet-B0 · ImageNet-1k | Top-1 accuracy: 42.2 -> 67.6 | Helped |
| Distill the student from a larger self-supervised teacher rather than train it directly with MoCo-V2 | MobileNet-V3-Large · ImageNet-1k | Top-1 accuracy: 36.3 -> 68.2 | Helped |
| Use a larger teacher | SEED with ResNet-18 student · ImageNet-1k | Top-1 accuracy: 52.5 -> 57.9 -> 62.7 | Helped |
| Replace SEED similarity-distribution matching with embedding L2 distance (ablation) | ResNet-18 distilled from ResNet-50 · ImageNet-1k | Top-1 accuracy: 57.9 -> 55.3 | Hurt |
| Replace SEED similarity-distribution matching with K-Means pseudo-labels (ablation) | ResNet-18 distilled from ResNet-50 · ImageNet-1k | Top-1 accuracy: 57.9 -> 51.0 | Hurt |
| Replace frozen-centroid K-Means with online clustering (ablation) | ResNet-18 distilled from ResNet-50 · ImageNet-1k | Top-1 accuracy: 51.0 -> 56.4 | Helped |
| Replace SEED similarity-distribution matching with binary contrastive loss (ablation) | ResNet-18 distilled from ResNet-50 · ImageNet-1k | Top-1 accuracy: 57.9 -> 57.4 | Hurt |
| Add the original MoCo-V2 SSL loss during SEED distillation (ablation) | ResNet-18 distilled from ResNet-50 · ImageNet-1k | Top-1 accuracy: 57.9 -> 57.6 | Hurt |
| Use cross-view rather than identical-view distillation (ablation) | ResNet-18 distilled from SWAV ResNet-50 · ImageNet-1k | Top-1 accuracy: 61.7 -> 58.2 | Hurt |
| Add six small patches while preserving identical-view distillation (ablation) | ResNet-18 distilled from SWAV ResNet-50 · ImageNet-1k | Top-1 accuracy: 61.7 -> 62.6 | Helped |
| Add six small patches with cross-view distillation (ablation) | ResNet-18 distilled from SWAV ResNet-50 · ImageNet-1k | Top-1 accuracy: 58.2 -> 61.9 | Helped |
| Use a deeper MoCo-V2 projection head (ablation) | MobileNet-V3-Large · ImageNet-1k linear evaluation | Top-1 accuracy: 19.0 -> 36.3 | Helped |
Closest papers by meaning
- Unsupervised Representation Transfer for Small Networks: I Believe I Can Distill On-the-Fly (NeurIPS 2021 · similarity 0.91)
- On the Efficacy of Small Self-Supervised Contrastive Models without Distillation Signals (AAAI 2022 · similarity 0.90)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.89)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.88)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.87)
- No Train, all Gain: Self-Supervised Gradients Improve Deep Frozen Representations (NeurIPS 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.