For modern DL, a large part also comes from regularization. And then also data augmentation. And self-supervision in whatever way, either prediction, masked pr

For modern DL, a large part also comes from regularization. And then also data augmentation. And self-supervision in whatever way, either prediction, masked prediction, contrastive losses, etc. Which all adds to the number of constraints / equations.