Research map / Vision-language and multimodal / Cross-modal learning
Dataset Growth
Ziheng Qin, zhaopan xu, YuKun Zhou, Kai Wang, Zangwei Zheng, Zebang Cheng, Hao Tang, Lei Shang, Baigui Sun, Radu Timofte, Xiaojiang Peng, Hongxun Yao and 1 more
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 MiniGPT-4 recaptioning to the pipeline (ablation) | InfoGrowth on BLIP, 0.4M CC3M · MSCOCO zero-shot retrieval | TR@1: 21.8 -> 34.1 | Improved |
| add gain-based sampling (ablation) | InfoGrowth on BLIP, 0.4M CC3M · MSCOCO zero-shot retrieval | TR@1: 21.8 -> 23.9 | Improved |
| combine recaptioning with gain-based sampling (ablation) | InfoGrowth · MSCOCO zero-shot retrieval | TR@1: 34.1 -> 37.1 | Improved |
| use k=4 neighbors instead of k=8 (arithmetic mean) (ablation) | gain calculator · MSCOCO zero-shot retrieval | TR@1: 33.6 -> 37.1 | Improved |
| use arithmetic instead of harmonic mean at k=4 (ablation) | gain calculator · MSCOCO zero-shot retrieval | TR@1: 35.3 -> 37.1 | Improved |
| use combined image and text gain instead of image gain only (ablation) | gain calculator · MSCOCO zero-shot retrieval | TR@1: 32.8 -> 37.1 | Improved |
| add alignment gain to info gain (ablation) | gain calculator · MSCOCO zero-shot retrieval | TR@1: 37.1 -> 33.7 | Hurt |
| process data with InfoGrowth under 25% shuffled captions | BLIP pretraining · MSCOCO zero-shot retrieval | TR@1: 15.0 -> 31.2 | Improved |
| InfoGrowth data (0.4M) instead of full CC3M (2.7M) | BLIP pretraining · MSCOCO zero-shot retrieval | TR@1: 34.7 -> 37.1 | Improved |
| InfoGrowth data (0.40M) instead of full CC3M on Flickr30K zero-shot | BLIP pretraining · Flickr30K zero-shot image-to-text | R@1: 87.4 -> 82.2 | Hurt |
| choose 0.67M selected data from 2.7M candidates rather than 1.35M (ablation) | InfoGrowth · MSCOCO zero-shot retrieval | TR@1: 41.1 -> 40.9 | No clear effect |
| InfoGrowth selection instead of UCB dynamic pruning | ResNet-50 · ImageNet-1K | top-1 accuracy (%): 75.3 -> 75.8 | Improved |
Closest papers by meaning
- Scaling Laws for Data Filtering-- Data Curation cannot be Compute Agnostic (CVPR 2024 · similarity 0.84)
- DataRater: Meta-Learned Dataset Curation (NeurIPS 2025 · similarity 0.82)
- Datasets, Documents, and Repetitions: The Practicalities of Unequal Data Quality (NeurIPS 2025 · similarity 0.82)
- A CLIP-Powered Framework for Robust and Generalizable Data Selection (ICLR 2025 · similarity 0.82)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.82)
- Data Pruning by Information Maximization (ICLR 2025 · similarity 0.82)
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.
