Efficiency Redefined: Impact of Reducing Data Acquisition Rate for Optimized TinyML in Resource-Constrained IoT Devices
Bidyut Saha, Riya Samanta, Soumya K. Ghosh, Ram Babu Roy · 2025
Tiny Machine Learning (TinyML) enables cost-effective, privacy-focused machine learning inference on micro-controller units (MCUs) connected to sensors. For these resource-limited settings, achieving an optimized design is essential. This study examines the impact of reducing data acquisition rates on TinyML models designed for time series classification, especially in battery-powered IoT devices with tight resource constraints. By decreasing sampling frequency, we aim to lower computational requirements—such as RAM usage, energy consumption, latency, and multiply-accumulate (MAC) operations—by approximately four times while preserving classification accuracy. Experiments conducted with six benchmark datasets (UCIHAR, WISDM, PAMAP2, MHEALTH, MIT-BIH, and PTB) indicate that reduced data acquisition rates substantially lessen energy demands and computational overhead with minimal impact on accuracy. For instance, lowering the acquisition rate by 75% for the MIT-BIH and PTB datasets resulted in a 60% reduction in RAM use, 75% fewer MAC operations, 74% shorter latency, and a 70% decrease in energy consumption, all without compromising accuracy. These findings offer key insights into the deployment of optimized TinyML models in constrained environments.