What are the pros and cons of learning an embedding via contrapositive loss vs through an autoencoder with a reconstruction loss.
We are guaranteeing correct class se
