NeurIPS 2020

Second Order Optimality in Decentralized Non-Convex Optimization via Perturbed Gradient Tracking

Meta Review

This paper gives a decentralized non-convex optimization algorithm that is guaranteed to find a second order stationary point. The paper deals with consensus issues that may arise in decentralized settings while maintaining interesting theoretical guarantees. The response also provided interesting experiment results. The paper would be stronger if it can improve the d dependency in phase II or explain why a dimension dependency is necessary.