Research map / Vision-language and multimodal / Cross-modal learning
Cross-Modal Knowledge Distillation without Paired Data: Theoretical Foundation and Algorithm
T. Khiem Tran, Anh Duc Chu, Quang Hung Pham, Phi Le Nguyen, Trong Nghia Hoang
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 bi-level optimization to joint feature and label alignment (ablation) | UCMKD · RAVDESS A→V | accuracy (%): 66.47 -> 73.83 | Improved |
| add bi-level optimization to joint feature and label alignment (ablation) | UCMKD · AVE A→V | accuracy (%): 28.11 -> 34.16 | Improved |
| add label alignment to feature alignment (ablation) | UCMKD · RAVDESS A→V | accuracy (%): 65.37 -> 73.83 | Improved |
| add feature alignment to label alignment (ablation) | UCMKD · AVE V→A | accuracy (%): 48.92 -> 52.24 | Improved |
| use l2 instead of l1 transport cost (ablation) | UCMKD · RAVDESS A→V | accuracy (%): 74.23 vs 73.83 | No clear effect |
| use l2 instead of angular transport cost (ablation) | UCMKD · CREMA-D V→A | accuracy (%): 62.41 -> 66.59 | Improved |
| use UCMKD unpaired instead of vanilla KD paired | UCMKD · AVE A→V | accuracy (%): 29.85 -> 34.16 | Improved |
| use UCMKD unpaired instead of vanilla KD paired | UCMKD · CREMA-D A→V | accuracy (%): 72.18 -> 71.64 | Hurt |
| use UCMKD instead of Cross Entropy | UCMKD · RAVDESS A→V (50% data) | accuracy (%): 59.87 -> 72.76 | Improved |
| use UCMKD instead of Cross Entropy | UCMKD · RAVDESS A→V (30% data) | accuracy (%): 54.51 -> 69.03 | Improved |
| use UCMKD instead of Cross Entropy with ViT backbone | UCMKD · RAVDESS A→V ViT-S | accuracy (%): 65.63 -> 80.32 | Improved |
| use UCMKD instead of C2KD in paired setting | UCMKD · RAVDESS V→A paired | accuracy (%): 82.78 -> 75.13 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- C2KD — C2KD: Bridging the Modality Gap for Cross-Modal Knowledge Distillation (CVPR 2024) · better on 0 of 1 shared result
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
- C2KD: Bridging the Modality Gap for Cross-Modal Knowledge Distillation (CVPR 2024 · similarity 0.94)
- Learning Modality Knowledge Alignment for Cross-Modality Transfer (ICML 2024 · similarity 0.88)
- On the Value of Cross-Modal Misalignment in Multimodal Representation Learning (NeurIPS 2025 · similarity 0.88)
- Align-KD: Distilling Cross-Modal Alignment Knowledge for Mobile Vision-Language Large Model Enhancement (CVPR 2025 · similarity 0.87)
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · similarity 0.87)
- When to Align, When to Predict: A Phase Diagram for Multimodal Learning (NeurIPS 2026 · 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.
