This paper by Arora, Ge, Neyshabur and Zhang proposes a compression based framework which purportedly explains the surprising generalization power of deep neural nets. The punchline is this - any neural network with certain robustness properties can be'compressed'. Compressed networks can be shown to generalize well, hence networks with these robustness properties are good candidates for networks that can hope to generalize well. The authors also show experimental evidence that these robustness properties are actually satisfied by real world neural nets. While I find the paper interesting, I am struggling with some of the technicalities.