This paper implements a practical, communication-efficient FL framework that combines update quantization with top-k sparsification with error feedback and evaluates its performance under realistic edge conditions, and provides reproducible MATLAB code and result artifacts.
Abstract
Edge IoT devices are increasingly targeted for on-device intelligence using federated learning (FL). However, conventional FL imposes heavy communication and energy costs that make it impractical for battery-constrained, bandwidth-limited deployments with heterogeneous (non-IID) data. In this paper we present a practical, communication-efficient FL framework that combines update quantization with top-k sparsification and evaluates its performance under realistic edge conditions. We implement the framework in MATLAB and conduct an empirical study on EMNIST and a synthetic IoT sensor dataset across varied Dirichlet non-IID severities, client dropout rates, and multiple random seeds. Our experiments show that combining low-bit quantization (4 bits) with sparsification (top 2–5%) yields large reductions in transmitted bytes and estimated uplink energy while maintaining useful model fidelity: Compared with uncompressed FedAvg, the proposed 4-bit quantization and top-1% sparsification with error feedback reduces cumulative communication by approximately 98.87%, while achieving a final accuracy of 70.30% ± 0.40% across three random seeds. We provide detailed convergence analysis, confidence intervals across seeds, and an energy model translating bytes to Joules to quantify device-level savings. Finally, we analyze failure modes and robustness under extreme heterogeneity and client dropout, and provide reproducible MATLAB code and result artifacts.
Experiments show that H-FedSN reduces communication costs by up to 477 times compared to baseline methods while maintaining high accuracy, making it well-suited for hierarchical FL in IoT deployments.
Jie-Chao Gao, Yuan-Gang Li, Jie Wang et al.· npj Artificial Intelligence· 2 citations
Experimental evaluation on MNIST, CIFAR-10, and Tiny ImageNet under a controlled federated simulation indicates that the proposed Federated CNN attains classification accuracy within 2.1 percentage points of a centralized baseline and reduces average per-round energy by approximately 43% relative to centralized trainin...
Ahmad Abu-Al-Aish, Tamer Shraa· Cluster Computing· 0 citations
Energy-Aware Adaptive Quantization and Freezing (EA-AQF), a unified framework that co-optimizes communication and computation, is presented, a unified framework that co-optimizes communication and computation and maintains robust convergence in highly heterogeneous tasks.
Experimental results on heterogeneous MNIST and CIFAR-10 settings show that QEF-GT-AdamW consistently improves robustness and convergence performance over representative DecL baselines while achieving favorable accuracy-communication trade-offs under limited wireless resources.
Thieu Van Nguyen, T. Nguyen, Ons Aouedi et al.· 0 citations
QuantFL is presented, a communication-efficient FL framework that starts from a pretrained initialisation and compresses per-layer client updates using bucket-based scalar quantisation, and is positioned as a lightweight update-compression primitive for scalable edge intelligence.
Charuka Herath, Y. Rahulamathavan, Varuna De-Silva et al.· IEEE International Symposium...· 0 citations
This paper proposes a communication-efficient adaptive federated learning algorithm for heterogeneous defect classification tasks that achieves competitive classification accuracy while reducing single-round training time by up to 70%.
Shuo He, He-Yang Wei, Congxian Bi et al.· Electronics· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 17, 2026
By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.
MIT News · Artificial Intelligence· news.mit.eduNov 17, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 9, 2026