SCAFFOLD
A classical optimization algorithm in federated learning that introduces control variates for variance reduction, effectively mitigating client drift. It provides well-established theoretical convergence guarantees for federated convex optimization problems.
Timeline (last 90 days)
Means of mitigating problems caused by non-IID data include the proximal term constraint of FedProx, the control variate correction of SCAFFOLD, and personalized federated learning.
Federated convex optimization algorithms such as SCAFFOLD and FedProx correct client drift by introducing variance reduction or proximal regularization terms, and have established near-optimal communication complexity bounds.
A key prerequisite for techniques such as SCAFFOLD and FedProx is that the operator F itself is a gradient field of some function, allowing the construction of a Lyapunov potential via function value descent to prove convergence, whereas a general monotone operator does not possess such a potential function.
All Facts (3)
Means of mitigating problems caused by non-IID data include the proximal term constraint of FedProx, the control variate correction of SCAFFOLD, and personalized federated learning.
50%UnverifiedFederated convex optimization algorithms such as SCAFFOLD and FedProx correct client drift by introducing variance reduction or proximal regularization terms, and have established near-optimal communication complexity bounds.
50%UnverifiedA key prerequisite for techniques such as SCAFFOLD and FedProx is that the operator F itself is a gradient field of some function, allowing the construction of a Lyapunov potential via function value descent to prove convergence, whereas a general monotone operator does not possess such a potential function.
50%