Reviews: Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks

I find the proposed technique interesting, but I believe that these problem might be better situated for a hardware venue. This seems to be more a contribution on hardware representations, as compared to previous techniques like xor-net which don't ask for a new hardware representation and had in mind an application to compression of networks for phones, and evaluated degradation of final performance. However, I do think that the paper could be of interest to the NIPS community as similar papers have been published in AI conferences. The paper is well written and easy to understand. The upsides of the method and the potential problems are well described and easy to follow.