Closed-Loop GAN for continual Learning