How would you handle a scenario where one modality dominates the loss during the training of a multimodal model?
