DART: Device-Native Adaptive Real-Time Training for Lifelong Learning on IoT Boards

Al-Ameen, S; Sudharsan, B; Alzacko, O; Al-Taie, R; Shah, T; Ranjan, R

doi:10.1109/BigData66926.2025.11401997

DART: Device-Native Adaptive Real-Time Training for Lifelong Learning on IoT Boards

Lookup NU author(s): Shamil Al-Ameen, Dr Tejal Shah, Professor Raj Ranjan

Downloads

Full text for this publication is not currently held within this repository. Alternative links are provided below where available.

Abstract

© 2025 IEEE. As deep learning models expand in scale and capability, their deployment on embedded and low-power platforms becomes increasingly challenging. The demand for continual adaptation, fast inference, and constrained-resource operation creates tension between performance and practicality. Traditional approaches to lifelong learning often impose prohibitive memory, compute, and latency costs, making them unsuitable for real-world edge devices. To address these challenges, we introduce DART, a resourceefficient lifelong learning framework tailored for embedded environments. DART integrates low-rank decomposition for memory compression, modulated gating for temporal adaptability, structured sparsity with support freezing for compute reduction, and MAS-based regularization to preserve past knowledge. We evaluate DART across benchmark datasets and IoT hardware platforms, demonstrating that it consistently achieves competitive accuracy while significantly reducing training and inference times. Specifically, DART reduces model size by up to 20× (down to 10-25 KB), accelerates training by 3-5×, and lowers inference latency by over 40% on devices such as the Raspberry Pi 5, while maintaining accuracy within 1-2% of state-of-the-art baselines. Our results highlight DART's ability to deliver scalable lifelong learning on constrained devices, enabling practical deployment of AI in dynamic real-world settings.