Reviews: Collaborative Deep Learning in Fixed Topology Networks

This paper explores a fixed peer-to-peer communication topology without parameter server. To demonstrate convergence, it shows that the Lyaounov functions that is minimized includes a regularizer term that incorporates the topology of the network. This leads to convergence rate bounds in the convex setting and convergence guarantees in the non-convex setting. This is original work of high technical quality, well positioned with a clear introduction. It is very rare to see proper convergence bounds in such a complex parallelization setting, the key to the proof is really neat (I did not check all the details).