Towards explainable meta-learning