cross and learn networks (cal)

An approach where two separate networks are trained simultaneously on the same task, and the gradients from each are used to update the other, potentially leading to better generalization.