Research map / Label-efficient and robust learning / Self-supervised and contrastive
An Empirical Study of Training Self-Supervised Vision Transformers
Xinlei Chen, Saining Xie, Kaiming He
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 |
|---|---|---|---|
| freeze the patch projection layer (random patch projection) (ablation) | MoCo v3 ViT-B/16 · ImageNet-1k linear probing (lr=1.5e-4) | top-1 accuracy: +1.7% to 73.4% | Improved |
| freeze the patch projection layer (random patch projection) (ablation) | SimCLR ViT-B/16 · ImageNet-1k linear probing | top-1 accuracy: +0.8% | Improved |
| freeze the patch projection layer (random patch projection) (ablation) | BYOL ViT-B/16 · ImageNet-1k linear probing | top-1 accuracy: +1.3% | Improved |
| freeze the patch projection layer (random patch projection) (ablation) | SwAV ViT · ImageNet-1k linear probing | top-1 accuracy: 65.8% -> 66.4% | Improved |
| increase batch size from 2k to 4k (ablation) | MoCo v3 ViT-B/16 · ImageNet-1k linear probing (100-ep) | top-1 accuracy: 72.6 -> 72.2 | Hurt |
| increase batch size from 4k to 6k (ablation) | MoCo v3 ViT-B/16 · ImageNet-1k linear probing (100-ep) | top-1 accuracy: 72.2 -> 69.7 | Hurt |
| increase base lr from 0.5e-4 to 1.0e-4 (ablation) | MoCo v3 ViT-B/16 · ImageNet-1k linear probing (100-ep, batch 4096) | top-1 accuracy: 70.4 -> 72.2 | Improved |
| add sin-cos position embedding (versus none) (ablation) | MoCo v3 ViT-B · ImageNet-1k linear probing (300-ep) | top-1 accuracy: 74.9 -> 76.5 | Improved |
| remove class token but keep the final LayerNorm (ablation) | MoCo v3 ViT-B · ImageNet-1k linear probing (300-ep) | top-1 accuracy: 76.5 -> 69.7 | Hurt |
| add BatchNorm in MLP heads (ablation) | MoCo v3 ViT-B · ImageNet-1k linear probing (300-ep) | top-1 accuracy: 74.4 -> 76.5 | Improved |
| add prediction MLP head (ablation) | MoCo v3 ViT-B · ImageNet-1k linear probing (300-ep) | top-1 accuracy: 75.5 -> 76.5 | Improved |
| add momentum encoder (m=0.99 instead of m=0) (ablation) | MoCo v3 ViT-B · ImageNet-1k linear probing (300-ep) | top-1 accuracy: 74.3 -> 76.5 | Improved |
Closest papers by meaning
- A Closer Look at Self-Supervised Lightweight Vision Transformers (ICML 2023 · similarity 0.94)
- DeiT III: Revenge of the ViT (ECCV 2022 · similarity 0.93)
- Efficient Self-supervised Vision Transformers for Representation Learning (ICLR 2022 · similarity 0.92)
- A Theoretical Analysis of Self-Supervised Learning for Vision Transformers (ICLR 2025 · similarity 0.91)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.91)
- Teaching Matters: Investigating the Role of Supervision in Vision Transformers (CVPR 2023 · similarity 0.91)
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.
