Gradually reducing the importance of parameters during training using differentiable masks rather than hard removal.