Meta-Learning Algorithms for Resource-Constrained Intelligent IoT Devices

Riandari, Fristi
Sihotang , Jonhariono

Jurnal Teknik Informatika C.I.T. Medicom
Vol 16 No 4 (2024): September: Intelligent Decision Support System (IDSS)

The rapid expansion of the Internet of Things (IoT) requires devices that can operate intelligently in dynamic environments despite severe hardware and energy constraints. Traditional machine learning models deployed on microcontroller-class IoT devices often struggle to adapt to new tasks, handle sensor noise, and maintain accuracy under changing environmental conditions. This research proposes a lightweight meta-learning framework specifically optimized for resource-constrained IoT platforms, combining gradient-based meta-learning techniques with model compression strategies such as quantization and pruning. The objective is to enable rapid few-shot adaptation, reduce computational overhead, and ensure robust performance in real-world IoT deployments. The study adopts a hardware-aware design approach, implementing the proposed model on ultra-low-power microcontrollers such as ARM Cortex-M series and ESP32. A two-phase training pipeline meta-training and on-device fine-tuning is used to evaluate adaptation speed, latency, memory footprint, accuracy, and energy consumption. Experimental results demonstrate that the lightweight meta-learning model adapts to new sensor-based tasks significantly faster than conventional supervised learning models while consuming substantially less energy. The model also shows improved resilience to environmental variations and sensor noise, outperforming baseline TinyML and standard meta-learning architectures under constrained conditions. Despite these promising results, the research identifies limitations related to computational cost, memory usage during adaptation, and the trade-off between model complexity and predictive accuracy. Nonetheless, the findings highlight the potential of meta-learning as a transformative approach for building intelligent, adaptive, and energy-efficient IoT systems. This study contributes to the advancement of TinyML and edge intelligence by providing a practical and scalable meta-learning solution tailored for ultra-low-power IoT devices.