Meta-learnt priors slow down catastrophic forgetting in neural networks