Multimodal interpretability

Orthogonality Regularization

Constraint applied during training to force the representations of different modalities in the shared latent space to be as independent as possible, avoiding redundancy and improving clarity of modality-specific explanations.

← Zurück