Research map / Vision-language and multimodal / Cross-modal learning
Understanding and Constructing Latent Modality Structures in Multi-Modal Representation Learning
Qian Jiang, Changyou Chen, Han Zhao, Liqun Chen, Qing Ping, Son Dinh Tran, 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 deep feature separation loss (orthogonal + contrastive + uniformity on independent features) | CLIP two-tower · Zero-shot CIFAR10/CIFAR100/ImageNet1K and linear probing | Top1 accuracy | Mixed |
| add Brownian-bridge loss | CLIP two-tower · CIFAR10 and CIFAR100 zero-shot | Top1 accuracy | Hurt |
| add Brownian-bridge loss | CLIP two-tower · ImageNet1K zero-shot and linear probing | Top1 accuracy: +3.73 zero-shot top1 vs CLIP; +1.79 linear average vs CLIP | Improved |
| add geometric consistency loss | CLIP two-tower · CIFAR10 and CIFAR100 zero-shot | Top1 accuracy: CIFAR10 +11.41; CIFAR100 +7.65 top1 vs CLIP | Improved |
| add geometric consistency loss | CLIP two-tower · Natural distribution shifts (ImageNetV2, Sketch, A, R) | Top1 accuracy | Improved |
| add deep feature separation, Brownian-bridge, or geometric consistency loss individually | CLIP two-tower · Linear probing average over 14 benchmarks | Top1 accuracy: Sep +5.68, Br +1.79, GC +2.96 vs CLIP | Improved |
| add Brownian-bridge loss | ALBEF fusion model · VQA test-dev | accuracy: +0.88 vs ALBEF | Improved |
| add deep feature separation loss | ALBEF fusion model · NLVR2 dev | accuracy: +0.69 vs ALBEF | Improved |
| add all three regularizers together (ablation) | ALBEF fusion model · VQA test-dev, NLVR2 dev, SNLI-VE test | accuracy: VQA +0.74 vs ALBEF; NLVR2 dev +1.82 vs ALBEF; SNLI-VE test +0.32 vs ALBEF | Improved |
| add deep feature separation, Brownian-bridge, or geometric consistency loss individually (ablation) | ALBEF fusion model · Flickr30K zero-shot text retrieval R@1 (COCO pretrain) | R@1: Sep +7.6, Br +7.0, GC +5.9 vs ALBEF | Improved |
| add Brownian-bridge loss (ablation) | ALBEF fusion model · Flickr30K zero-shot image retrieval R@1 (COCO pretrain) | R@1: +6.3 vs ALBEF; +1.8 vs CODIS | Improved |
| add modality-gap-reducing alignment loss (1/<Z_T,Z_V>^2) (ablation) | ALBEF-style pretraining on COCO · Zero-shot Flickr30K retrieval | retrieval performance | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CyCLIP — CyCLIP: Cyclic Contrastive Language-Image Pretraining (NeurIPS 2022) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Aligning Multimodal Representations through an Information Bottleneck (ICML 2025 · similarity 0.91)
- Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion (ICLR 2025 · similarity 0.90)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.89)
- Guiding Cross-Modal Representations with MLLM Priors via Preference Alignment (NeurIPS 2025 · similarity 0.89)
- Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (NeurIPS 2026 · similarity 0.89)
- Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.89)
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.
