Researchers have developed a joint entropy-aware quantization method for Tiny Machine Learning (TinyML) models on low-power IoT devices. The goal is to improve energy efficiency and accuracy in human activity recognition (HAR) at the network edge, where computational resources are limited. This approach allows models to operate with fewer bits per parameter, reducing memory and power consumption without significantly sacrificing performance.

The method addresses a key challenge in TinyML implementation: how to compress complex models to run effectively on microcontrollers with very restricted resources. Traditional quantization often reduces accuracy, whereas this new approach optimizes bit allocation based on the information importance of each model layer, measured through joint entropy. This results in a more compact and efficient representation of neural network weights and activations.

The results show that this quantization scheme achieves an optimal balance between energy efficiency and HAR accuracy. Compared to existing methods, a considerable reduction in power consumption and memory usage is observed, while maintaining a high activity recognition capability. This is crucial for applications such as wearables, health monitoring, and home automation, where battery life and processing capacity are critical factors.