Multimodal Learning

1. Which unit must stay whole when splitting multimodal data?
2. What does the control where image–text pairs are randomly shuffled verify?
3. Why remove one modality during evaluation?
4. What behavior must be defined when a sensor is missing in production?
5. Does a higher global performance prove that all modalities are useful?

Answers. 1. The entity that connects the views — scene, person, session or document — so that another view of the same case does not cross partitions. 2. That the gain depends on a real alignment rather than on unimodal shortcuts or class frequency. 3. An ablation measures the contribution of each source and reveals a decorative modality. 4. An explicit degraded mode, abstention or tested fallback. 5. No: ablations, contradictory cases, missing inputs and comparison to the best unimodal model at the same budget are required.