Why might a multimodal training objective weight different modality losses separately?

1. To balance their contribution to optimization
2. To prove the model will be fair automatically
3. To guarantee every modality has identical raw values
4. To remove the need for paired-data quality checks