Adaptive Edge-Device Collaborative Framework for Image Classification

Deep neural network (DNN) has emerged as a superior technology in mobile applications. However, due to the computation-intensive nature of DNN’s execution, it is challenging for mobile devices with limited computational power and energy to deploy them and meet real-time requirements. This paper proposes an adaptive edge-device collaborative (AEC) framework for DNN inference optimization. Different from previous works, our approach dynamically optimizes model segmentation and feature compression based on the current environmental state, ensuring that accuracy requirements are met. Specifically, the problem is divided into two subproblems: segmentation and feature compression. A policy-based deep reinforcement learning (DRL) model determines the optimal segmentation layer, while an attention-based compression algorithm maximizes feature compression. We conducted the experiments in an edge computing scenario consisting of an NVIDIA Jetson Nano and an edge server equipped with RTX 3090. Experiments show that the scheme significantly optimizes DNN inference in time-varying network environments. The proposed adaptive dynamic segmentation outperforms the local scheme, edge-only scheme, and the static split computing scheme, which can effectively reduce the latency and energy consumption while satisfying the accuracy requirements with varying network environments. Based on the measured results, the inference performance of AEC for different network environments is improved by more than 30% on average compared to all the baseline algorithms in this paper.