Uncertainty-based Continual Learning with Adaptive Regularization