Adversarial Training with Orthogonal Regularization
2020
A regularization objective that penalizes deviation from an orthogonality constraint among selected vectors or transformations.
The penalty depends on which learned quantities are compared and on the norm used to measure their deviation from orthogonality.