Optimizing Stochastic Gradient Push under Broadcast Communications

Name: Optimizing Stochastic Gradient Push under Broadcast Communications
Uploaded: 2026-10-15T08:45:23.030362+00:00
Duration: 755 s
Description: The authors developed a method to optimize mixing matrices for Stochastic Gradient Push (SGP) in decentralized federated learning under broadcast communications. They derived an objective function based on graph-theoretic parameters, leading to an efficient design algorithm. This approach reduces convergence time by 11–45% compared to existing methods without sacrificing model quality.

The authors developed a method to optimize mixing matrices for Stochastic Gradient Push (SGP) in decentralized federated learning under broadcast communications. They derived an objective function based on graph-theoretic parameters, leading to an efficient design algorithm. This approach reduces convergence time by 11–45% compared to existing methods without sacrificing model quality.

Open the full episode