Enhanced AI as a Service at the Edge via Transformer Network

TL;DR

The paper addresses energy-efficient DNN inference offloading in mobile edge computing by modeling each task as a DAG and formulating a combinatorial optimization to minimize device energy under latency and capacity constraints. It introduces a transformer-based DNN to learn offloading policies from historical problem instances, producing near-optimal solutions in polynomial time. The dataset focuses on common architectures such as MobileNetV1, ResNet18, and VGG16, trained with Adam on a penalized objective. Experimental results show energy reductions up to 18% per completed task and lower task failure rates in resource-constrained MEC settings, highlighting the practical potential for scalable AIaaS at the edge.

Abstract

Artificial intelligence (AI) has become a pivotal force in reshaping next generation mobile networks. Edge computing holds promise in enabling AI as a service (AIaaS) for prompt decision-making by offloading deep neural network (DNN) inference tasks to the edge. However, current methodologies exhibit limitations in efficiently offloading the tasks, leading to possible resource underutilization and waste of mobile devices' energy. To tackle these issues, in this paper, we study AIaaS at the edge and propose an efficient offloading mechanism for renowned DNN architectures like ResNet and VGG16. We model the inference tasks as directed acyclic graphs and formulate a problem that aims to minimize the devices' energy consumption while adhering to their latency requirements and accounting for servers' capacity. To effectively solve this problem, we utilize a transformer DNN architecture. By training on historical data, we obtain a feasible and near-optimal solution to the problem. Our findings reveal that the proposed transformer model improves energy efficiency compared to established baseline schemes. Notably, when edge computing resources are limited, our model exhibits an 18\% reduction in energy consumption and significantly decreases task failure compared to existing works.

Figures (6)

• Figure 1: An instance of the MEC system with edge AI.
• Figure 2: DNN task of a mobile device, showcasing the layer dependency.
• Figure 3: Proposed transformer architecture for task offloading problem.
• Figure 4: Average energy consumption for completing a task for different algorithms under various DNN models. To ensure a fair comparison, the energy consumed by failed tasks is excluded.
• Figure 5: Energy consumption across different transmission rates for various DNN tasks: a) ResNet18, b) MobileNetV1, c) VGG16.