Research map / Vision-language and multimodal / CLIP and zero-shot
SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency
Yangyang Guo, Mohan Kankanhalli
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 |
|---|---|---|---|
| prune both redundant and ill-matched samples rather than redundant only (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 33.51 -> 35.51 | Improved |
| prune both redundant and ill-matched samples rather than ill-matched only (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 35.19 -> 35.51 | Improved |
| prune ill-matched samples alone instead of redundant alone (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 33.51 -> 35.19 | Improved |
| use mutation epoch 3 instead of 4 (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 33.80 -> 35.51 | Improved |
| use mutation epoch 3 instead of 2 (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 36.28 -> 35.51 | Mixed |
| use 20:20 redundant to ill-matched ratio instead of 30:10 (ablation) | SCAN · ImageNet zero-shot (ViT-S/16) | ZS@5: 34.84 -> 35.51 | Improved |
| increase pruning ratio (ablation) | SCAN · CLIP models | downstream accuracy | Hurt |
| use dynamic SCAN pruning instead of static SemDeDup coreset | SCAN · ImageNet zero-shot (ViT-B/32) | ZS@1: 07.20 -> 16.48 | Improved |
| use SCAN static coreset instead of SemDeDup | SCAN · ImageNet zero-shot (ViT-B/32) | ZS@1: 07.20 -> 13.16 | Improved |
| use SCAN (30% pruning) instead of full CLIP data on RN101 | SCAN · ImageNet fine-tune (CC3M+) | top-1: 75.39 -> 74.61 | No clear effect |
| use SCAN instead of CLIP full data on Swin-Base | SCAN · ImageNet zero-shot (CC3M+) | top-1: 13.98 -> 17.50 | Improved |
| use SCAN instead of Info-Batch on MoCo ViT-B/16 | SCAN · ImageNet | top-1: 78.46 -> 79.15 | Improved |
Closest papers by meaning
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.87)
- Filter Images First, Generate Instructions Later: Pre-Instruction Data Selection for Visual Instruction Tuning (CVPR 2025 · similarity 0.87)
- AmorLIP: Efficient Language-Image Pretraining via Amortization (NeurIPS 2025 · similarity 0.87)
- DELT: A Simple Diversity-driven EarlyLate Training for Dataset Distillation (CVPR 2025 · similarity 0.87)
- Accelerating Augmentation Invariance Pretraining (NeurIPS 2024 · similarity 0.86)
- Demystifying CLIP Data (ICLR 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.
