Learning paradigm that pulls related samples together and pushes unrelated ones apart in embedding space.