Year: 2024
Type: preprint
Abstract: In bandwidth-constrained federated learning~(FL) settings, the repeated upload of high-dimensional model updates from agents to a central server constitutes the primary bottleneck, often rendering standard FL infeasible within practical communication budgets. We propose \emph{FedScalar}, a communication-efficient FL algorithm in which each agent uploads only two scalar values per round, regardless of the model dimension~$d$. Each agent encodes its local update difference as an inner product with a locally generated random vector and transmits the resulting scalar together with the generating seed, enabling the server to reconstruct an unbiased gradient estimate without any high-dimensional transmission. We prove that \emph{FedScalar} achieves a convergence rate of $O(d/\sqrt{K})$ to a stationary point for smooth non-convex loss functions, and show that adopting a Rademacher distribution for the random vector reduces the aggregation variance compared to the Gaussian case. Numerical simulations confirm that the dimension-free upload cost translates into significant improvements in wall-clock time and energy efficiency over \emph{FedAvg} and \emph{QSGD} in bandwidth-constrained settings.
Source: arXiv (Cornell University)
Authors: Rostami, M., Kia, S. S.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4403882782