Research map / Vision-language and multimodal / Cross-modal learning
Learning Unseen Modality Interaction
Yunhua Zhang, Hazel Doughty, Cees G. M. Snoek
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 feature projection into a common space (ablation) | vanilla multimodal transformer · Video classification | top-1 accuracy (%): 17.2 -> 21.3 | Improved |
| Add dual branch prediction with pseudo-supervision (ablation) | transformer with feature projection · Video classification | top-1 accuracy (%): 21.3 -> 23.7 | Improved |
| Use vanilla transformer instead of simple late fusion averaging (ablation) | multimodal fusion · Video classification | top-1 accuracy (%): 18.0 -> 17.2 | Hurt |
| Add feature alignment loss L_align (ablation) | full model · Video classification | top-1 accuracy (%): 20.2 -> 21.4 | Improved |
| Add pseudo-supervised loss L_pseudo (ablation) | full model with L_align · Video classification | top-1 accuracy (%): 21.4 -> 23.7 | Improved |
| Reduce parameters from 14.4M to 7.2M to match vanilla transformer (ablation) | full model · EPIC-Kitchens | top-1 accuracy (%): 23.7 -> 21.4 | Mixed |
| Replace feature projection with average pool plus linear layer (ablation) | full model · EPIC-Kitchens RGB+Flow+Audio | top-1 accuracy (%): 23.7 -> 19.0 | Hurt |
| Add noise N(0,1) to one modality, compare to vanilla transformer | full model · Video classification | top-1 accuracy (%): 10.1 -> 18.0 | Improved |
| Vary feature token length k* (ablation) | feature projection · Video classification | top-1 accuracy (%): 20.9 -> 23.7 -> 21.2 | Mixed |
| Vary number of learnable tokens n_u from 32 to 128 (ablation) | feature alignment loss · Multimedia retrieval | MnR: 73.2 -> 66.2 | Improved |
| Change pseudo to supervised token split from 50:50 to 30:70 or 70:30 (ablation) | dual branch prediction · Video classification | top-1 accuracy (%): 23.7 -> 22.9 | Hurt |
| Vary alpha from 0 to 3000 (ablation) | pseudo loss weight · Video classification | top-1 accuracy (%): 21.4 -> 23.7 | Improved |
Closest papers by meaning
- What to align in multimodal contrastive learning? (ICLR 2025 · similarity 0.90)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.89)
- Enhancing Multimodal Cooperation via Sample-level Modality Valuation (CVPR 2024 · similarity 0.89)
- Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.89)
- ReconBoost: Boosting Can Achieve Modality Reconcilement (ICML 2024 · similarity 0.89)
- Multimodal Learning Without Labeled Multimodal Data: Guarantees and Applications (ICLR 2024 · 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.
