Research map / Video / Video-language understanding
Scaling Multimodal Pre-Training via Cross-Modality Gradient Harmonization
Junru Wu, Yi Liang, feng han, Hassan Akbari, Zhangyang Wang, Cong Yu
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 gradient realignment and gradient-based curriculum learning | VATT-MA-Medium pretrained on HowTo100M + AudioSet + YouTube8M · UCF101 | Top1: 88.28 -> 89.70 | Improved |
| add gradient realignment and gradient-based curriculum learning | VATT-MA-Medium pretrained on HowTo100M + AudioSet + YouTube8M · HMDB51 | Top1: 65.84 -> 64.35 | Hurt |
| add gradient realignment and gradient-based curriculum learning | VATT-MA-Medium pretrained on HowTo100M + AudioSet + YouTube8M · MSR-VTT | Rank: 56.00 -> 45.00 | Improved |
| add gradient realignment and gradient-based curriculum learning | VATT-MA-Medium pretrained on HowTo100M + AudioSet + YouTube8M · ESC50 | Top1: 80.75 -> 84.50 | Improved |
| scale up the video-audio gradient and scale down video-text gradient | VATT-MA-Medium · YouCook2 | Rank: 93.50 -> 168.90 | Hurt |
| apply cross-modality gradient realignment | VATT-MA-Medium pretrained on HowTo100M · YouCook2 | Rank: 93.50 -> 47.00 | Improved |
| apply gradient-based curriculum learning | VATT-MA-Medium pretrained on HowTo100M · YouCook2 | Rank: 93.50 -> 40.00 | Improved |
| combine gradient realignment with curriculum learning | VATT-MA-Medium pretrained on HowTo100M · YouCook2 | R@10: 17.10 -> 25.87 | Improved |
| use a progressively increasing gamma schedule from -0.3 to 0.0 (ablation) | gradient-based curriculum learning on VATT · UCF101 | Top1: 10.05 -> 79.10 | Improved |
| raise the starting gamma from -0.3 to -0.2 (ablation) | gradient-based curriculum learning on VATT · UCF101 | Top1: 79.10 -> 77.75 | Hurt |
| raise the starting gamma from -0.3 to -0.1 (ablation) | gradient-based curriculum learning on VATT · UCF101 | Top1: 79.10 -> 44.49 | Hurt |
| use fixed gamma 0.0 instead of the progressive schedule (ablation) | gradient-based curriculum learning on VATT · UCF101 | Top1: 79.10 -> 10.05 | Hurt |
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.
- VATT — VATT: Transformers for Multimodal Self-Supervised Learning from Raw Video, Audio and Text (NeurIPS 2021) · better on 9 of 12 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
- Understanding Multimodal Learning From Modality Fusion and Alignment Perspectives (TPAMI 2026 · similarity 0.90)
- When to Align, When to Predict: A Phase Diagram for Multimodal Learning (NeurIPS 2026 · similarity 0.88)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.88)
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · similarity 0.87)
- Improving Multimodal Learning Balance and Sufficiency through Data Remixing (ICML 2025 · similarity 0.87)
- Enhancing Multimodal Cooperation via Sample-level Modality Valuation (CVPR 2024 · 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.
