The Upsides of Turbulence: Baselining Gossip Learning in Dynamic Settings

In dynamic settings, fully distributed gossip-based learning schemes have recently gained interest due to their better scalability, robustness, and enhanced privacy protection compared to server-based architectures. However, existing approaches to their performance characterization either assume stable connectivity among nodes or are ad-hoc for specific trace-based mobility patterns. Thus, in dynamic settings, there is currently a poor understanding of the conditions under which gossip-based learning schemes are feasible, and of their main performance tradeoffs. In this work, we start addressing this issue by performing a first baselining of Gossip Learning (GL) on random Time-Varying Graphs (TVG), to get a first-order characterization of their main performance patterns in dynamic settings. The use of random TVG enables a fine-grained and accurate characterization of GL effectiveness as a function of the main system parameters while abstracting from scenario-specific features of patterns of communication and mobility (e.g., induced by road grids or measured mobility traces). Our results suggest that GL schemes are robust to node mobility and comparable in accuracy and convergence speed to Federated Learning architectures, over a wide range of operational conditions. We show that the final model accuracy is robust against data dispersion across nodes as well as against very low rates of exchanges across nodes.