Worth noting that all reviewers have voted reject: https://openreview.net/forum?id=MGWsPGogLH

But it would be nice to see an implementation of it.

> peers who were not as familiar with the literature on how computational complexity relates to machine learning. As a result, we have over-optimized for educating that audience, as opposed to stating our own results clearly and compellingly.

I wish this version of the paper could be shared somewhere, what's wrong with making a subject accessible?

This is huge! Can't wait to see some implementations.