Research map / Vision-language and multimodal / Cross-modal learning
Multi-Modal Alignment Using Representation Codebook
Jiali Duan, Liqun Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, Trishul Chilimbi
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 intra-modal alignment to cross-modal alignment (ablation) | CODIS without codebook · MSCOCO (5K), zero-shot text retrieval | R@1: 68.60 -> 69.86 | Improved |
| Add codebook loss using teacher features to cross-modal alignment only (ablation) | CODIS cross-alignment-only baseline · MSCOCO (5K), zero-shot text retrieval | R@1: 68.60 -> 70.74 | Improved |
| Add codebook loss using teacher features to intra- and cross-modal alignment (ablation) | CODIS · MSCOCO (5K), zero-shot text retrieval | R@1: 69.86 -> 71.10 | Improved |
| Use teacher rather than student features to compute codebook loss (ablation) | CODIS with intra- and cross-modal alignment · MSCOCO (5K), zero-shot text retrieval | R@1: 71.12 -> 71.10 | No clear effect |
| Add intra-modal alignment (ablation) | CODIS zero-shot retrieval · MSCOCO (5K), zero-shot image retrieval | R@1: 50.10 -> 50.52 | Improved |
| Add intra-modal alignment (ablation) | CODIS zero-shot retrieval · Flickr30K (1K), zero-shot text retrieval | R@1: 84.90 -> 85.80 | Improved |
| Add intra-modal alignment (ablation) | CODIS zero-shot retrieval · Flickr30K (1K), zero-shot image retrieval | R@1: 68.18 -> 69.70 | Improved |
| Use 500 codewords instead of the default codebook size (ablation) | CODIS pretrained on MSCOCO · MSCOCO zero-shot text retrieval | TR@1: 59.38 -> 55.52 | Hurt |
| Vary codebook size to 0.5x (ablation) | CODIS pretrained on MSCOCO · MSCOCO zero-shot text retrieval | TR@1: 59.38 -> 58.66 | Hurt |
| Vary codebook size to 2.0x (ablation) | CODIS pretrained on MSCOCO · MSCOCO zero-shot text retrieval | TR@1: 59.38 -> 59.02 | No clear effect |
| Vary codebook size to 3K codewords (ablation) | CODIS pretrained on MSCOCO · MSCOCO zero-shot text retrieval | TR@1: 59.38 -> 58.96 | No clear effect |
| Use codebook-based cluster alignment with teacher-student distillation | CODIS · MSCOCO (5K), zero-shot text retrieval | R@1: 68.6 -> 71.5 | Improved |
Closest papers by meaning
- Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment (NeurIPS 2024 · similarity 0.89)
- Do Vision and Language Encoders Represent the World Similarly? (CVPR 2024 · similarity 0.89)
- Linguistic-Aware Patch Slimming Framework for Fine-grained Cross-Modal Alignment (CVPR 2024 · similarity 0.89)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · similarity 0.89)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.88)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.88)
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.
