Comparing Model Compression Techniques for MCU-Based Federated TinyML

This paper focuses on combining Tiny Machine Learning (TinyML) and Federated Learning (FL) to enable continuous and distributed model training directly on resource-constrained IoT-edge devices like microcontrollers. The study discusses the challenges of training global models in such environments, focusing on communication overhead and energy consumption. Optimizing communication efficiency is crucial for the scalable and large deployment of FL, especially in IoT and edge settings with limited bandwidth and intermittent connections. To address these issues, we propose a two-stage distributed approach that employs model compression techniques encompassing pruning, quantization, and gradient sparsification to reduce data transmission and computation costs. We also provide an implementation in C extending the AIfES framework, optimized for resource-limited devices, facilitating real-time model compression and efficient FL operations on embedded tiny hardware.