Metaheuristic-Driven Dynamic Split Point Determination for Efficient IoT-Based Deep Learning

Abstract

The growing incorporation of Artificial Intelligence (AI) inside the Internet of Things (IoT) requires effective deployment of Deep Neural Networks (DNN) on resource-limited devices. Nonetheless, bandwidth constraints and computing burdens provide considerable obstacles, especially in real-time applications. Collaborative Intelligence (CI) tackles this issue by allocating DNN inference between edge and cloud contexts; nevertheless, identifying the ideal division point is a significant barrier for improving efficiency. This study introduces Dynamic Split Point Determination with Meta-Heuristic Algorithms (DSP-MHs), an innovative framework that adaptively determines optimal split points utilizing Grey Wolf Optimization (DSP-GWO), Particle Swarm Optimization (DSP-PSO), Artificial Bee Colony (DSP-ABC), and Artificial Rabbit Optimization (DSP-ARO). In contrast to current static or heuristic-based methods, MH-based methods improve inference time, energy consumption, and memory use dynamically in real-time situations. Experimental findings indicate that DSP-MH diminishes inference time by as much as 99.86%, reduces energy consumption by 99.85%, and enhances memory utilization by 99.98%, markedly surpassing conventional techniques. The suggested system adjusts to fluctuating network circumstances and computing demands, rendering it exceptionally appropriate for AI-driven IoT applications. This study illustrates that metaheuristic optimization significantly enhances the efficiency of distributed DNN inference, providing a scalable and adaptive solution for future IoT systems.