Reviews: Estimating Learnability in the Sublinear Data Regime

When I started reading the paper, I actually liked it. But when I got to page 5, I was very disappointed and annoyed. It might not be intentional, but it felt almost deceptive that the authors did not mention very relevant prior art till then and made it look like many of the previous literature as their own contribution. Authors claim that they propose an algorithm that uses O(sqrt(d)) samples for estimating the variance of noise. In fact, after a read of this paper and previous papers, what I got was: O(sqrt(d)) was essentially known before (for most cases) and new results are in fact much weaker.