Co-Learning for missing arbitrary modalities in multi-modal classification

Preprint: abs/2607.24683

This work addresses multi-modal classification under arbitrary missing-modality conditions, where any subset of modalities may be unavailable at inference time due to operational constraints. Rather than relying on robust fusion, the proposed framework adopts multi-modal co-learning to transfer complementary information between modalities during training. Two approaches are introduced, exploiting inter-modal collaboration at the feature and decision levels. Experiments on two multi-modal classification benchmarks demonstrate substantial robustness improvements across different missing-modality scenarios. The results show complementary behavior between the methods: one is particularly effective when a single modality is missing, while the other performs better under extreme conditions where only one modality remains available.

Recommended citation: Mena, Francisco, et al. "Co-Learning for Missing Arbitrary Modalities in Multi-modal Classification." arXiv preprint arXiv:2607.24683 (2026).
Download Paper