Research map / Vision-language and multimodal / Cross-modal learning
Multimodal Contrastive Training for Visual Representation Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| pre-train with multimodal contrastive objectives rather than image-only MoCo-v2 | ResNet-50 visual backbone · ImageNet-1K val | top-1 accuracy (%): 49.3 -> 54.9 | Helped |
| add dataset tag supervision (ablation) | multimodal contrastive pre-training · ImageNet-1K val | top-1 accuracy (%): 54.9 -> 55.3 | Helped |
| add tag loss J_tag to image contrastive loss J_ii (ablation) | COCO image pre-training · ImageNet-1K val | top-1 accuracy (%): 49.3% -> 50.2% | Helped |
| use POS tagging and noun filtering instead of the dataset tags (ablation) | tag-supervised multimodal pre-training · ImageNet-1K val | top-1 accuracy (%): 55.3% to 54.5 % | Hurt |
| use distinct MLP heads rather than shared MLP heads (ablation) | multimodal visual representation learning · ImageNet-1K val | top-1 accuracy (%): 49.6%, 50.2% and 52.3% vs. 53.6%, 53.8% and 53.9% | Helped |
| use 128-d intra-modal and 1024-d inter-modal representations (ablation) | distinct-head multimodal framework · ImageNet-1K val | top-1 accuracy (%): 53.6%, 53.8% and 53.9% -> 54.9% | Helped |
| train the text encoder from scratch instead of using pre-trained BERT (ablation) | Ours · ImageNet-1K val | top-1 accuracy (%): 54.9 -> 54.6 | Hurt |
| train with all five losses rather than only J_ii and J_tag (ablation) | multimodal contrastive model · ImageNet-1K val | top-1 accuracy (%): 50.2% -> 55.3% | Helped |
| use multimodal pre-training rather than inter-modal losses alone (ablation) | COCO image and text encoders without downstream fine-tuning · COCO 1K test-set, image-to-text | R@1: 13.2 -> 24.1 | Helped |
| use multimodal pre-training rather than inter-modal losses alone (ablation) | COCO image and text encoders without downstream fine-tuning · COCO 1K test-set, text-to-image | R@1: 9.5 -> 18.6 | Helped |
| pre-train with multimodal contrastive learning rather than Stock tagging supervision | ResNet-50 visual backbone · Stock 50K test-set | mIOU@5: 12.86% -> 13.81% | Helped |
Closest papers by meaning
- ITO: Multi-View Alignment and Training-Time Fusion for Image-Text Pretraining (NeurIPS 2026 · similarity 0.90)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.87)
- Aligning Multimodal Representations through an Information Bottleneck (ICML 2025 · similarity 0.87)
- Continual Multimodal Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- What to align in multimodal contrastive learning? (ICLR 2025 · similarity 0.87)
- Context-Aware Multimodal Pretraining (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.