Exploring intra-task relations to improve meta-learning algorithms