Research map / Vision-language and multimodal / Cross-modal learning
Decoupling Common and Unique Representations for Multimodal Self-supervised Learning
Yi Wang, Conrad M Albrecht, Nassim Ait Ali Braham, Chenying Liu, Zhitong Xiong, Xiao Xiang Zhu
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 |
|---|---|---|---|
| add intra-modal training and modality decoupling to cross-modal Barlow Twins (ablation) | DeCUR · GeoNRW RGB-DEM | mIoU: 33.6 -> 36.9 | Improved |
| add modality decoupling alone without intra-modal training (ablation) | DeCUR · SAR-optical and RGB-DEM | mAP / mIoU: 80.3 -> 80.1 (SAR-optical); 33.6 -> 34.3 (RGB-DEM) | Mixed |
| add intra-modal training on top of decoupling (ablation) | DeCUR · SAR-optical | mAP: 80.1 -> 81.7 | Improved |
| add decoupling on top of intra-modal training (ablation) | DeCUR · GeoNRW RGB-DEM | mIoU: 35.2 -> 36.9 | Improved |
| add deformable attention without residual connection (ablation) | DeCUR · label-limited frozen settings | mAP: -0.1 to -1.2 at 1% labels | Hurt |
| add residual deformable attention (RDA) (ablation) | DeCUR · BigEarthNet and GeoNRW frozen | mAP / mIoU: +0.4 to +2.4 in 7 of 8 settings | Improved |
| use cross-modal deformable attention instead of RDA (ablation) | DeCUR · BigEarthNet-SAR 1% | mAP: 74.4 -> 72.5 | Hurt |
| replace RDA with CBAM (ablation) | DeCUR · BigEarthNet-SAR 1% | mAP: 74.4 -> 73.1 | Hurt |
| decouple up to 50% of embedding dimensions as unique (ablation) | DeCUR · SAR-optical and RGB-DEM | downstream performance | No clear effect |
| swap cross-modal SimCLR for DeCUR with ViT-S/16 | ViT backbone · BigEarthNet-SAR 1% frozen | mAP: 73.9 -> 75.8 | Improved |
| pretrain with DeCUR instead of single-modal Barlow Twins | SAR-only transfer · BigEarthNet SAR-only 1% labels | mAP: 71.2/73.3 -> 74.4/76.0 | Improved |
Closest papers by meaning
- What to align in multimodal contrastive learning? (ICLR 2025 · similarity 0.86)
- Unity by Diversity: Improved Representation Learning for Multimodal VAEs (NeurIPS 2024 · similarity 0.85)
- Learning Shared Representations from Unpaired Data (NeurIPS 2025 · similarity 0.85)
- Aligning Multimodal Representations through an Information Bottleneck (ICML 2025 · similarity 0.85)
- A Closer Look at Multimodal Representation Collapse (ICML 2025 · similarity 0.85)
- RedCore: Relative Advantage Aware Cross-Modal Representation Learning for Missing Modalities with Imbalanced Missing Rates (AAAI 2024 · 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.
