Local Extra SGD
An extragradient stochastic gradient descent algorithm for federated learning settings, in which each client performs multiple local Extra SGD iterations before aggregation. It introduces a prediction step to stabilize variational inequality solving, balancing communication efficiency and convergence.
Timeline (last 90 days)
The paper proves that the classical Local Extra SGD algorithm, through more refined analysis, can achieve tighter convergence guarantees than previously known for general smooth and monotone variational inequalities.
In the federated variant of Local Extra SGD, each client performs multiple rounds of Extra SGD iterations locally before synchronizing model parameters or gradient information to a central server for aggregation.
All Facts (2)
The paper proves that the classical Local Extra SGD algorithm, through more refined analysis, can achieve tighter convergence guarantees than previously known for general smooth and monotone variational inequalities.
50%UnverifiedIn the federated variant of Local Extra SGD, each client performs multiple rounds of Extra SGD iterations locally before synchronizing model parameters or gradient information to a central server for aggregation.
50%