What is contrastive learning in the context of multimodal deep learning?

Option D captures the general, task-agnostic definition of contrastive learning: given pairs of inputs labeled as similar (positive pairs) or dissimilar (negative pairs), the training objective pulls positive pairs' representations closer together in embedding space while pushing negative pairs' representations further apart — typically implemented via losses like InfoNCE, triplet loss, or contrastive loss with a margin. This is the mechanism underlying self-supervised representation learning broadly, not only in multimodal settings.

Option E correctly applies this general principle to the multimodal case: for the *same* object described across modalities (e.g., an image of a dog and the caption "a dog"), the model should increase representational similarity, since they refer to the same underlying entity; for *different* objects across modalities (an image of a dog paired with the caption "a cat"), the model should decrease similarity. This is exactly CLIP's training objective, tested elsewhere in this set — matching image-text pairs pulled together, mismatched pairs pushed apart.

Options B and C both invert this relationship — B increases similarity for *different* objects and decreases it for *same* objects, and C similarly reverses the correct direction — describing the opposite of what contrastive learning is designed to achieve, making both clearly incorrect distractors that test careful reading of directionality. Option A is too vague and mischaracterizes contrastive learning as a generative/manipulation technique rather than a representation-learning objective.