A technique that reverses the gradient updates made during training to remove learned information about specific data.