thinks we might have left out some relevant works on this topic, by citing [37] as well as some learning theory

We thank all the reviewers for the time reading our paper! W e plan to include that if the paper gets in. We're afraid R2 has misunderstood our work. Our main contribution is to show "can Neu ral Net works learn some Generally, our work is not about "there exist learning methods Our main contribution is to show that "Neural Networks can be better learners than kernels (esp. This was not known at all in the distribution-free setting, which we have emphasized in the introduction. For [37], its separation works only when the solution (i.e., the network weights) corresponds to a "max-margin This is not necessarily efficiently learnable.