For a task like image captioning, the model learns relationships between image embeddings and text embeddings.