Agent-based simulation of distributed edge learning for deep neural networks

Training deep neural networks on edge devices reduces latency and preserves privacy, but limited compute and bandwidth require a trade-off between accuracy and processing time. This paper uses the deep learning multi-processing agent-based simulation framework to model distributed deep learning coordination under heterogeneous edge conditions and quantify the trade-off between communication cost and accuracy. Deep learning multi-processing is a lightweight simulator in which each agent simulates an edge node with distinct computational power for exchanging model parameters. To simulate the network dynamics of edge learning systems, deep learning multi-processing models different communication topologies among the machine learning models hosted on edge nodes participating in federated learning, focusing on how communication strategies for exchanging model parameters affect distributed training under resource constraints. Simulating the network aspect of distributed deep learning offers a cost-effective way to evaluate training performance under realistic edge conditions during the design phase of deep learning edge-enabled systems. The findings of such simulation are valuable for industry sectors developing intelligent systems for smart cities, including video-based traffic control, surveillance analytics and sensor-driven urban monitoring, where efficient and adaptive deep learning parallelization is required at the network edge.