Design of lightweight neural networks for resource-constrained devices

Abstract

Practical realization of modern artificial intelligence systems, especially deep neural networks, on edge platforms presents a daunting challenge. The root cause lies in the critical gap between the computational requirements of these models and the drastically limited capabilities of edge platforms in terms of processing power, memory, storage, and energy consumption. This constraint often requires applications to rely on cloud processing, which presents essential problems in terms of added latency, privacy, and persistent internet connectivity. To overcome this problem, this study presents an architecture for designing and deploying compact neural networks. The methodology begins with the choice of the MobileNet architecture as an initial reference, then adopts advanced model compression schemes, i.e., pruning to eliminate redundant neural connections, and quantization to reduce the numerical precision of weights, substantially contributing to model size reduction and computational requirements. The optimized models were then implemented and evaluated on Raspberry Pi and Arduino Nano boards to test their usability in practical situations. Experimental results clearly demonstrate that optimized models realized an energy consumption reduction of 40% and a latency reduction of 43.75%, while retaining an impressive level of performance in terms of an accuracy loss of less than 3%. This research provides evidence in support of bridging the gap between complex AI and resource-limited hardware, thereby enabling the realization of real-time, compact, and secure on-device intelligent applications.