深度学习模型终端环境自适应方法研究

Context-aware adaptation of deep learning models for IoT devices

随着人工智能和物联网的快速发展与融合,智能物联网AIoT正成长为一个极具前景的新兴前沿领域,其中深度学习模型的终端运行是其主要特征之一.针对智能物联网应用场景动态多样,以及物联网终端(智能手机、可穿戴及其他嵌入式设备等)计算和存储资源受限等问题,深度学习模型环境自适应正成为一种新的模型演化方式.其旨在确保适当性能的条件下,能自适应地根据环境变化动态调整模型,从而降低资源消耗、提高运算效率.具体来说,它需要主动感知环境、任务性能需求和平台资源约束等动态需求,进而通过终端模型的自适应压缩、云边端模型分割、领域自适应等方法,实现深度学习模型对终端环境的动态自适应和持续演化.本文围绕深度学习模型自适应问题,从其概念、系统架构、研究挑战与关键技术等不同方面进行阐述和讨论,并介绍我们在这方面的研究实践.

The rapid development of both Artificial Intelligence(AI)and the Internet of Things(IoT),has cultivated the new research area:the Artificial Intelligence of Things(AIoT).AIoT is used to deploy many different deep learning models on a variety of local IoT terminals including smartphones,wearables,and other embedded devices.Adapting to these dynamic and varied AIoT application scenarios,and the IoT platform resources(e.g.,computation and storage resources)available in each diverse,requires a novel scheme for improving on device environmental adaptability.Deep learning models aim to dynamically adjust either the model structure,the calculation scheme,or both,of them specifically to adapt to the environment context.They must reduce costs and improve computational efficiency while creating negligible performance degradation.Specifically,an environmental adaptation evolution framework must actively and continuously assess the constantly changing environmental context including factors,such as application data,knowledge base,task-related performance requirements,and platform-imposed resource constraints.Then it must adopt on-demand model compression,model segmentation,and domain adaptation techniques to achieve a appropriate balance between the model’s performance and the environment’s budget.This paper focuses on making deep learning models for context-aware adaptation.We discuss the system architecture and core technologies solving this problem requires.We address research challenges in this area,and introduce our pilot research practice in this field.