Joint entropy-aware quantization for energy-efficient TinyML human activity recognition on IoT edge devices
- 1IRIS Team, MIASI Laboratory, Faculty of Applied Sciences, Nador, Mohammed Premier University
Abstract
Sensor-based human activity recognition (HAR) is a core task in wearable healthcare, assisted living, sports monitoring, rehabilitation, and context-aware Internet-of-Things (IoT) systems. However, deploying HAR models on microcontroller-class edge devices remains challenging because continuous inference must satisfy strict constraints on Flash memory, SRAM usage, latency, and energy consumption. Conventional TinyML-HAR pipelines commonly rely on uniform INT8 or INT4 quantization, which reduces computational cost but assumes that all sensor channels, neural layers, activations, and activity windows require the same numerical precision. This assumption is suboptimal because wearable motion signals exhibit heterogeneous information content: static activities often produce low-entropy inertial patterns, whereas dynamic and transitional activities generate richer and more quantization-sensitive signal distributions. To address this limitation, this paper proposes JEAQ-HAR, a joint entropy-aware quantization framework for energy-efficient TinyML HAR on IoT edge devices. JEAQ-HAR estimates entropy at the sensor-channel, neural-layer, activation, and input-window levels to guide a hierarchical mixed-precision allocation policy across the HAR inference pipeline. High-entropy sensor axes, quantization-sensitive layers, and complex motion windows are assigned higher precision, while low-entropy and redundant components are quantized more aggressively. The framework also introduces an entropy-distortion regularization objective to preserve activation-distribution structure during quantization-aware training and an adaptive inference policy that selects low-bit or higher-precision execution paths according to window-level motion complexity. Experiments on UCI-HAR, WISDM, and PAMAP2 show that JEAQ-HAR achieves macro-F1 scores of 95.1%, 93.1%, and 90.1%, respectively, with an average gap of only 0.47 percentage points from the FP32 backbone. Across the three datasets, the average serialized deployment package size is reduced from 159.6 KB for FP32 to 29.17 KB for JEAQ-HAR, corresponding to an 81.7% reduction relative to the FP32 serialized package, rather than merely the raw trainable-weight tensor. On microcontroller-class devices, JEAQ-HAR reduces the complete measured end-to-end inference latency and energy from 14.0 ms and 1.07 mJ for FP32 to 5.93 ms and 0.47 mJ, including entropy computation, routing, packed-precision handling, quantized inference, and decision generation. These gains represent end-to-end workload reductions rather than speedups from native INT6 arithmetic. Compared with uniform INT4 quantization, JEAQ-HAR improves average macro-F1 by 4.8 percentage points while maintaining a compact memory footprint and low inference cost. The ablation study confirms that sensor-level entropy, layer-level entropy, window-level entropy, entropy-distortion regularization, and adaptive inference each contribute to recognition stability or deployment efficiency. These results show that combining entropy-guided precision allocation with deployment-aware TinyML evaluation provides an effective and practical strategy for accurate, compact, and energy-efficient HAR on constrained IoT edge devices.
Related
- Edge efficient human activity recognition using a quantized patch based transformer
- Symbolic transfer entropy reveals enhanced cardiac autonomic influence on cortical activity in infants with congenital heart disease
- An efficient brain tumor classification system with accurate recognition and complementary downsampling blocks
- Dual-stream attention and feature fusion for human activity recognition via wearable sensors
- GlucoSense: early diabetes risk assessment via fusion of human activity recognition and glucose prediction.
BibTeX
@article{lamaakal2026joint,
title = {{Joint entropy-aware quantization for energy-efficient TinyML human activity recognition on IoT edge devices}},
author = {Ismail Lamaakal and Chaymae Yahyati and Khalid El Makkaoui and Ibrahim Ouahbi},
journal = {Scientific Reports},
year = {2026},
month = sep,
doi = {10.1038/s41598-026-72683-9},
url = {https://doi.org/10.1038/s41598-026-72683-9},
note = {CC BY},
}
RIS
TY - JOUR
TI - Joint entropy-aware quantization for energy-efficient TinyML human activity recognition on IoT edge devices
AU - Ismail Lamaakal
AU - Chaymae Yahyati
AU - Khalid El Makkaoui
AU - Ibrahim Ouahbi
PY - 2026
DA - 2026/09/01
JO - Scientific Reports
DO - 10.1038/s41598-026-72683-9
UR - https://doi.org/10.1038/s41598-026-72683-9
AB - Sensor-based human activity recognition (HAR) is a core task in wearable healthcare, assisted living, sports monitoring, rehabilitation, and context-aware Internet-of-Things (IoT) systems. However, deploying HAR models on microcontroller-class edge devices remains challenging because continuous inference must satisfy strict constraints on Flash memory, SRAM usage, latency, and energy consumption. Conventional TinyML-HAR pipelines commonly rely on uniform INT8 or INT4 quantization, which reduces computational cost but assumes that all sensor channels, neural layers, activations, and activity windows require the same numerical precision. This assumption is suboptimal because wearable motion signals exhibit heterogeneous information content: static activities often produce low-entropy inertial patterns, whereas dynamic and transitional activities generate richer and more quantization-sensitive signal distributions. To address this limitation, this paper proposes JEAQ-HAR, a joint entropy-aware quantization framework for energy-efficient TinyML HAR on IoT edge devices. JEAQ-HAR estimates entropy at the sensor-channel, neural-layer, activation, and input-window levels to guide a hierarchical mixed-precision allocation policy across the HAR inference pipeline. High-entropy sensor axes, quantization-sensitive layers, and complex motion windows are assigned higher precision, while low-entropy and redundant components are quantized more aggressively. The framework also introduces an entropy-distortion regularization objective to preserve activation-distribution structure during quantization-aware training and an adaptive inference policy that selects low-bit or higher-precision execution paths according to window-level motion complexity. Experiments on UCI-HAR, WISDM, and PAMAP2 show that JEAQ-HAR achieves macro-F1 scores of 95.1%, 93.1%, and 90.1%, respectively, with an average gap of only 0.47 percentage points from the FP32 backbone. Across the three datasets, the average serialized deployment package size is reduced from 159.6 KB for FP32 to 29.17 KB for JEAQ-HAR, corresponding to an 81.7% reduction relative to the FP32 serialized package, rather than merely the raw trainable-weight tensor. On microcontroller-class devices, JEAQ-HAR reduces the complete measured end-to-end inference latency and energy from 14.0 ms and 1.07 mJ for FP32 to 5.93 ms and 0.47 mJ, including entropy computation, routing, packed-precision handling, quantized inference, and decision generation. These gains represent end-to-end workload reductions rather than speedups from native INT6 arithmetic. Compared with uniform INT4 quantization, JEAQ-HAR improves average macro-F1 by 4.8 percentage points while maintaining a compact memory footprint and low inference cost. The ablation study confirms that sensor-level entropy, layer-level entropy, window-level entropy, entropy-distortion regularization, and adaptive inference each contribute to recognition stability or deployment efficiency. These results show that combining entropy-guided precision allocation with deployment-aware TinyML evaluation provides an effective and practical strategy for accurate, compact, and energy-efficient HAR on constrained IoT edge devices.
SN - 2045-2322
ER -