Research map / Label-efficient and robust learning / Domain adaptation and generalization
CHIPS: Efficient CLIP Adaptation via Curvature-aware Hybrid Influence-based Data Selection
Xinlin Zhuang, Yichen Li, Xiwei Liu, Haolin Yang, Yifan Lu, Ziyun Zou, Yulong Li, Huifa Li, Dongliang Chen, Qinglei Wang, Weiyang Liu, Ying Qian and 2 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 |
|---|---|---|---|
| select CPT data with CHIPS instead of random selection | MetaCLIP-B16-400M CPT on BIOMEDICA · Medical avg (17 tasks), r=10% | avg accuracy: 24.78 -> 27.03 | Improved |
| use CHIPS instead of TracIn for selection | MetaCLIP-B16-400M CPT · Medical avg, r=10/20/30% | avg accuracy: 26.46 -> 27.03; 26.63 -> 28.20; 25.68 -> 29.96 | Improved |
| CHIPS 30% subset instead of full-dataset CPT | MetaCLIP-B16-400M CPT · Medical avg | avg accuracy: 31.51 -> 29.96 | Hurt |
| CHIPS selection versus random on general-domain retention | MetaCLIP-B16-400M CPT · General CLS, r=10% | classification accuracy: 52.21 (Random) -> 47.88 (CHIPS) | Hurt |
| CHIPS instead of TracIn across seven backbones/scales | MetaCLIP B32/B16/L14/H14 (400M and CC) · Medical CLS, 10% kept | accuracy: e.g. L14-400M 27.08 -> 29.73 | Improved |
| add learnability margin term to alignment score (ablation) | CHIPS · Medical avg, r=30% | avg accuracy: 27.84 -> 28.50 | Improved |
| add learnability margin term to alignment score (ablation) | CHIPS · Medical avg, r=10% | avg accuracy: 25.98 -> 25.95 | No clear effect |
| add domain relevance weight on top of alignment+margin (ablation) | CHIPS · Medical avg, r=10/20/30% | avg accuracy: 25.95 -> 27.03; 27.92 -> 28.20; 28.50 -> 29.96 | Improved |
| restrict scoring subspace to text projection only versus all end-point params (ablation) | CHIPS · Medical avg, r=10% | avg accuracy: 27.03 -> 26.95 | No clear effect |
| score using only logit/temperature parameter (ablation) | CHIPS · Medical avg, r=10% | avg accuracy: 27.03 -> 23.83 | Hurt |
| increase target eval set size from 50 to 200 samples per task (ablation) | CHIPS · Medical avg, r=10% | avg accuracy: 26.25 -> 27.03 | Improved |
| mix negative-pair curvature at moderate alpha (0.6-0.8) vs extremes (ablation) | CHIPS · Medical avg, r=10% | avg accuracy: 25.66 (alpha 0.4) / 26.07 (alpha 1.0) -> 27.05 (alpha 0.6) | Improved |
Closest papers by meaning
- CLIPLoss and Norm-Based Data Selection Methods for Multimodal Contrastive Learning (NeurIPS 2024 · similarity 0.87)
- A CLIP-Powered Framework for Robust and Generalizable Data Selection (ICLR 2025 · similarity 0.87)
- LP++: A Surprisingly Strong Linear Probe for Few-Shot CLIP (CVPR 2024 · similarity 0.86)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.85)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.85)
- Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling (ICLR 2025 · similarity 0.85)
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.
