Adding vs. Averaging in Distributed Primal-Dual Optimization

Distributed optimization methods for large-scale machine learning suffer from a communication bottleneck. It is difficult to reduce this bottleneck while still efficiently and accurately aggregating p