Federated Learning Under Importance Sampling

Federated learning encapsulates distributed strategies that are managed by a central unit. Since it relies on using selected number of agents at each iteration, and since agent, in turn, taps into its local data, is only natural to study optimal sampling policies for selecting their data federated implementations. Usually, uniform schemes used. However, this work, we examine the effect importance devise non-uniformly guided performance measure. We find involving without replacement, resulting architecture controlled two factors related variability model across agents. illustrate theoretical findings with experiments simulated real show improvement results from proposed strategies.