An Energy-Aware Approach to Design Self-Adaptive AI-based Applications on the Edge

Alessandro Tundo,Marco Mobilio,Shashikant Ilager,Ivona Brandić,Ezio Bartocci,Leonardo Mariani

2023-08-31

Abstract:The advent of edge devices dedicated to machine learning tasks enabled the execution of AI-based applications that efficiently process and classify the data acquired by the resource-constrained devices populating the Internet of Things. The proliferation of such applications (e.g., critical monitoring in smart cities) demands new strategies to make these systems also sustainable from an energetic point of view. In this paper, we present an energy-aware approach for the design and deployment of self-adaptive AI-based applications that can balance application objectives (e.g., accuracy in object detection and frames processing rate) with energy consumption. We address the problem of determining the set of configurations that can be used to self-adapt the system with a meta-heuristic search procedure that only needs a small number of empirical samples. The final set of configurations are selected using weighted gray relational analysis, and mapped to the operation modes of the self-adaptive application. We validate our approach on an AI-based application for pedestrian detection. Results show that our self-adaptive application can outperform non-adaptive baseline configurations by saving up to 81\% of energy while loosing only between 2% and 6% in accuracy.

Software Engineering,Machine Learning,Systems and Control

What problem does this paper attempt to address?

The paper aims to address the balance between energy consumption and application performance (such as detection accuracy and frame processing rate) for AI applications on edge devices. Specifically, with the widespread use of IoT devices in scenarios like smart cities, these devices need to process large amounts of data and are often limited by battery capacity or renewable energy supply. Therefore, how to reduce energy consumption while ensuring application performance has become a key challenge.

The authors propose an energy-aware approach to design adaptive AI applications. This method can dynamically adjust system configurations under different runtime conditions to achieve a balance between energy saving and maintaining application goals (such as detection accuracy and frame processing speed). Through a meta-heuristic search process combined with Weighted Grey Relational Analysis (WGRA), researchers can filter out a set of configurations suitable for different operating modes from a vast configuration space. Experimental results show that this adaptive approach can save up to 81% of energy consumption compared to non-adaptive baseline configurations while maintaining a certain level of accuracy (with a maximum loss of 2%-6% accuracy) and improve processing speed by 77%-233%.

Overall, the main contribution of the paper is the proposal of a novel energy-aware framework that enables developers to build energy-efficient and high-performance AI edge applications, especially in resource-intensive application scenarios such as pedestrian detection systems.