OnlineStructuredMeta-learning-Appendix

A.1 HomogeneousDataset Follow the traditional meta-learning setting [1, 4, 5], each knowledge block is comprised of a convolutionlayer,abatchnormalization layer,andaReLUactivation. In this section, we analyze task learning efficiency on the homogeneous dataset (i.e., Rainbow MNIST). Specifically,we follow [2]to analyze the amount of data needed to learn each task and show the results in Figure 4.