Co-Learning for Missing Arbitrary Modalities in Multi-modal Classification

Researchers proposed a co-learning framework for multimodal classification that does not rely on fixed missing-modality patterns.

The approach combines feature-level and decision-level information to stay robust when any subset of modalities is absent.

On two benchmark datasets, it showed stronger performance under both mild and extreme missing-modality conditions.

The work tackles a common real-world challenge in multimodal AI, such as sensor failures or privacy-driven input loss.