AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Differential-informed Sample Selection Accelerates Multimodal Contrastive Learning

Zihua Zhao, Feng Hong, Mengxi Chen, Pengyi Chen, Benyuan Liu, Jiangchao Yao, Ya Zhang, Yanfeng Wang

ICCV 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
select samples by differential of CLIPScore (DISSect-Warmup) instead of random online selectionCLIP-ResNet50 on CC3M · MS-COCO 5K zero-shot, 30% selectionIR@1: 19.36 -> 21.34Improved
select by differential instead of small-loss trickCLIP-ResNet50 on CC3M · MS-COCO 5K zero-shot, 50% selectionIR@1: 20.05 -> 23.58Improved
use temporal ensembling instead of warm-up for the historical score (ablation)DISSect · MS-COCO 5K zero-shot, 50% selectionIR@1: 23.58 -> 22.60No clear effect
use temporal ensembling instead of warm-up for the historical score (ablation)DISSect · Flickr30K 1K zero-shot, 50% selectionIR@1: 39.40 -> 42.70Mixed
add DISSect selection to CLIP pretraining on YFCC15MCLIP-ResNet50 · MS-COCO 5K zero-shot, 50% selectionR@Sum: 162.43 -> 167.79Improved
add DISSect selection to CLIP pretraining on CC12MCLIP-ResNet50 · MS-COCO 5K zero-shot, 50% selectionR@Sum: 178.79 -> 187.61Improved
replace SCAN with DISSectBLIP-Base · NLVR2 devaccuracy: 72.19 -> 73.99Improved
replace SCAN with DISSectCLIP-ResNet101 · MS-COCO 5K zero-shot, 50% selectionR@Sum: 143.52 -> 150.18Improved
use single-network DISSect instead of dual-network GSCCLIP-ResNet50 on CC3M · MS-COCO 5K zero-shotR@Sum: 144.72 -> 148.44Improved
use loss-based Big Loss selection instead of random onlineCLIP-ResNet50 on CC3M · MS-COCO 5K zero-shot, 30% selectionIR@1: 19.36 -> 15.54Hurt
use oracle-free DISSect instead of oracle CLIPScoreCLIP-ResNet50 on CC3M · MS-COCO 5K zero-shot, 70% selectionIR@1: 23.74 -> 22.74Hurt
apply DISSect to reach full-data performanceCLIP-ResNet50 on CC3M · MS-COCO retrievaltraining iterations: approximately 2.85x accelerationImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy