Research map / Label-efficient and robust learning / Recognition and long-tail
NeuCLIP: Efficient Large-Scale CLIP Training with Neural Normalizer Optimization
Xiyuan Wei, Chih-Jen Lin, Tianbao Yang
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 |
|---|---|---|---|
| replace OpenCLIP/FastCLIP with NeuCLIP | CLIP training · DFN-14M | Datacomp Average: 37.78 -> 39.16 (OpenCLIP), 38.45 -> 39.16 (FastCLIP) | Improved |
| replace OpenCLIP with NeuCLIP at largest scale | CLIP training · DFN-1B | Datacomp Average: 56.25 -> 57.34 | Improved |
| use a unified objective instead of separate objectives for CLIP and NPN (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 38.63 -> 39.16 | Improved |
| use single-layer inductive-bias NPN instead of MLP (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 38.58 -> 39.16 | Improved |
| re-initialize NPNs every 500 iterations instead of never (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 38.48 -> 39.16 | Improved |
| restart too rarely (Tr 500 -> 2,500) (ablation) | NeuCLIP · CC12M | Datacomp Average: 31.89 -> 31.47 | Hurt |
| restart too often (Tr 500 -> 20) (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 39.16 -> 38.41 | Hurt |
| increase NPN updates per iteration from 1 to 10 (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 39.02 -> 39.16 | Improved |
| increase NPN updates beyond 10 (10 -> 50) (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 39.16 -> 38.03 | Hurt |
| increase number of prototypes m from 4,096 to 8,192 (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 39.16 -> 39.25 | No clear effect |
| increase number of prototypes m from 1,024 to 4,096 (ablation) | NeuCLIP · DFN-14M | Datacomp Average: 38.57 -> 39.16 | Improved |
| add NPN training to OpenCLIP-style training | NeuCLIP · ViT-B/16 | time overhead: 5.98% overhead | No clear effect |
Closest papers by meaning
- AmorLIP: Efficient Language-Image Pretraining via Amortization (NeurIPS 2025 · similarity 0.91)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.88)
- Meta CLIP 2: A Worldwide Scaling Recipe (NeurIPS 2025 · similarity 0.87)
- CLIPLoss and Norm-Based Data Selection Methods for Multimodal Contrastive Learning (NeurIPS 2024 · similarity 0.87)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.87)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.87)
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.
