Research map / Vision-language and multimodal / Cross-modal learning
Contrastive Vision-Language Pre-training with Limited Resources
Quan Cui, Boyan Zhou, Yu Guo, Weidong Yin, Hao Wu, Osamu Yoshie, Yubo Chen
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 debiased sampling (same-dataset batches) (ablation) | naive baseline · MS-COCO 5K zero-shot | RSUM: 371.6 -> 392.3 | Improved |
| add debiased sampling (same-dataset batches) (ablation) | naive baseline · Flickr30K 1K zero-shot | RSUM: 483.3 -> 510.1 | Improved |
| add coin flipping mixup (ablation) | baseline + debiased sampling · MS-COCO 5K zero-shot | RSUM: 392.3 -> 402.8 | Improved |
| add coin flipping mixup (ablation) | baseline + debiased sampling · Flickr30K 1K zero-shot | RSUM: 510.1 -> 519.2 | Improved |
| add gradient reserved gather (ablation) | baseline + data methods · MS-COCO 5K zero-shot | RSUM: 402.8 -> 415.0 | Improved |
| add gradient reserved gather (ablation) | baseline + data methods · Flickr30K 1K zero-shot | RSUM: 519.2 -> 524.1 | Improved |
| increase batch size from 1,024 to 16,384 (ablation) | ZeroVL · MS-COCO 5K zero-shot | RSUM: 415.0 -> 425.0 | Improved |
| increase batch size from 1,024 to 16,384 (ablation) | ZeroVL · Flickr30K 1K zero-shot | RSUM: 524.1 -> 536.2 | Improved |
| use decoupled gradient accumulation (16 steps) on 8 GPUs instead of batch 16,384 on 128 GPUs (ablation) | ZeroVL · MS-COCO 5K zero-shot | RSUM: 425.0 vs 425.2 | No clear effect |
| add identical random seed for the two forwards in DGA (ablation) | decoupled gradient accumulation · MS-COCO 5K zero-shot | RSUM: 413.4 -> 425.2 | Improved |
| add identical random seed for the two forwards in DGA (ablation) | decoupled gradient accumulation · Flickr30K 1K zero-shot | RSUM: 527.1 -> 535.9 | Improved |
| add TokenDrop (mask 25% of ViT tokens) (ablation) | ZeroVL with DGA · MS-COCO 5K zero-shot | RSUM / GPU time: 425.2 -> 424.8; ~680 -> ~470 hr | No clear effect |
Closest papers by meaning
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.91)
- FFF: Fixing Flawed Foundations in Contrastive Pre-Training Results in Very Strong Vision-Language Models (CVPR 2024 · similarity 0.90)
- Extending Multi-modal Contrastive Representations (NeurIPS 2024 · similarity 0.89)
- ITO: Multi-View Alignment and Training-Time Fusion for Image-Text Pretraining (NeurIPS 2026 · similarity 0.89)
- Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment (NeurIPS 2024 · similarity 0.89)
- MLIP: Efficient Multi-Perspective Language-Image Pretraining with Exhaustive Data Utilization (ICML 2024 · similarity 0.89)
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.
