Loss function that trains embeddings by comparing an anchor example to a positive and negative example simultaneously.