Weaknesses: The primary weaknesses of the approach is that the authors do not sufficiently convince the reader that the knowledge block approach is necessary to yield the achieved performance improvements. Generally, it is unclear if the machinery of the knowledge blocks truly result in modular knowledge representations. While the authors present a framework that achieves reasonably strong performance on their chosen benchmarks, I think the paper does not effectively present a contribution that authors could build on in future work. I believe ablations of the model would help convince a reader of the utility of the knowledge block formulation. In particular, an experiment with the same architecture but without the meta-knowledge pathway construction would help convince a reader that this step is necessary.

This paper convinced the reviewers that the modular approach to meta-learning proposed therein outperforms rivals in a series of experiments. There were concerns about novelty, as the approach can be seen as a combination of prior approaches, but honestly I don't see why this should be seen as a negative. It is not the duty of scientific progress to surprise us with regard to the sources from where it is drawn. If combining existing approaches exploits synergies and complementarities, then a paper demonstrating the empirical success of such a combination is worth sharing with the community.

Learning quickly is of great importance for machine intelligence deployed in online platforms. With the capability of transferring knowledge from learned tasks, meta-learning has shown its effectiveness in online scenarios by continuously updating the model with the learned prior. However, current online meta-learning algorithms are limited to learn a globally-shared meta-learner, which may lead to sub-optimal results when the tasks contain heterogeneous information that are difficult to share. We overcome this limitation by proposing an online structured meta-learning (OSML) framework. Inspired by the knowledge organization of human and hierarchical feature representation, OSML explicitly disentangles the meta-learner as a meta-hierarchical graph with different knowledge blocks.