Accelerating Convergence in Split Learning for Time-Varying and Resource-Limited Environments

Split Learning (SL) is a distributed learning paradigm characterized by the neural network partitioning into distinct client-side and server-side segments. This work focuses on optimizing SL's performance for time-varying and resource-constrained systems. Specifically, the main goal in the paper is to determine the optimal cut layer position in the neural network that minimizes the total training delay, i.e. maximizes the rate of convergence. The performance analysis shows that the proposed cut layer selection algorithm outperforms State of the Art solutions.