Deep Edge Intelligence Architecture Key Features Enabling Technologies and Challenges Prabath Abeysekara1 Hai Dong1 and A. K. Qin2

2025-05-06 0 0 1.49MB 15 页 10玖币

侵权投诉

Deep Edge Intelligence: Architecture, Key

Features, Enabling Technologies and Challenges

Prabath Abeysekara1, Hai Dong1, and A. K. Qin2

1School of Computing Technologies, RMIT University, Melbourne, Australia

{prabath.abeysekara, hai.dong}@rmit.edu.au

2Computer Science and Software Engineering, Swinburne University of Technology,

Hawthorn, Australia

kqin@swin.edu.au

Abstract. With the breakthroughs in Deep Learning, recent years have

witnessed a massive surge in Artiﬁcial Intelligence applications and ser-

vices. Meanwhile, the rapid advances in Mobile Computing and Internet

of Things has also given rise to billions of mobile and smart sensing

devices connected to the Internet, generating zettabytes of data at the

network edge. The opportunity to combine these two domains of tech-

nologies to power interconnected devices with intelligence is likely to

pave the way for a new wave of technology revolutions. Embracing this

technology revolution, in this article, we present a novel computing vi-

sion named Deep Edge Intelligence (DEI). DEI employs Deep Learning,

Artiﬁcial Intelligence, Cloud and Edge Computing, 5G/6G networks, In-

ternet of Things, Microservices, etc. aiming to provision reliable and

secure intelligence services to every person and organisation at any place

with better user experience. The vision, system architecture, key layers

and features of DEI are also detailed. Finally, we reveal the key enabling

technologies and research challenges associated with it.

Keywords: Deep Edge Intelligence ·Artiﬁcial Intelligence ·Deep Learn-

ing ·Edge Computing ·Internet of Things.

1 Introduction

The pervasive adaptation of Deep Learning (DL) driven by the recent advance-

ments in Mobile Computing (MC) and Artiﬁcial Intelligence (AI) has opened

up a myriad of opportunities across several application domains. Examples for

such applications include autonomous cars, video analytics, and cognitive assis-

tance technologies. [1][2]. These applications powered by billions of mobile and

Internet of Things (IoT) devices connected to the internet generate zillion bytes

of data at the network edge. The opportunity to combine these two domains of

technologies to power interconnected devices with intelligence is likely to pave

the way for a new wave of technology revolutions. More than the technologies

themselves, what lies at the forefront of this technological revolution is the vision

to make the lives of people better and eﬃcient. To realise the aforementioned vi-

sion, this data generated in exorbitant volumes at rapid velocities at the edge of

arXiv:2210.12944v1 [cs.DC] 24 Oct 2022

2 Prabath Abeysekara, Hai Dong, and A. K. Qin

the network needs to be processed and analysed to derive solutions to everyday

problems of people based on DL and other AI techniques.

DL techniques in particular, having demonstrated unparalleled performance

primarily in Computer Vision and Natural Language Processing ﬁelds, are often

subjected to signiﬁcant computation and memory costs as well as massive data

requirements. This poses a great challenge to empower devices at the network

edge with advanced analytics based on DL capabilities [3]. In such a setting,

the emerging Edge Computing (EC) paradigm provides a promising way to en-

able this. Sitting in close proximity to end-users and services, EC aims to pro-

vide computing, storage and network resources for applications. In other words,

leveraging the distributed computing concepts to push computational loads from

the network core to the network edge with the aim to provide faster responses

to end users, EC provides an enticing platform to enable DL-based intelligent

applications at the edge of the network.

Recent attempts to combine EC and AI thereby fully unleashing the potential

values of big data generated at the edge has led to the Edge Intelligence (EI)

paradigm. EI brings together EC and AI together to shift intelligence to the edge,

relieving the network infrastructure with exponentially increasing network stress.

In the process, it also promises end-users with context-aware, faster intelligent

services at the edge of the network. Although there has been a rapidly growing

amount of research focused on enabling DL-based EI in the recent past, there

is an urgent need for a more holistic framework that drives such approaches to

enable true edge-native DL strategies that push the DL frontier to the network

edge more comprehensively. We envision that such a holistic framework will bring

together multiple key technologies to come up with a comprehensive strategy

that facilitates and makes recommendations for autonomic, deeply-integrated

and environment-aware, privacy-preserving, collaborative and trustworthy DL

applications in an EC environment. We also hope that such an approach will

bring researchers and enterprises together to outline a new class of AI strategies,

algorithms and a collection of reference architectures as well as applications

addressing the challenges of DL applications in the aforesaid setting.

To realise the aforementioned vision, we present a novel computing framework

named Deep Edge Intelligence (DEI), which is a combination of DL, AI, EC and

AIoT. DEI enables the development and deployment of DL and AI techniques,

based on EC, on edge devices, e.g., AIoT devices, where the data is generated,

aiming to provide AI for every person and every organisation at any place.

The remainder of this paper is structured as follows. Section 2 provides a

comprehensive survey of existing EI approaches. Section 3 introduces the vision

of DEI and also the systems architecture of DEI. Meanwhile, Section 4 com-

prehensively discusses the key enabling technologies of DEI. Section 5 presents

possible research challenges as well as opportunities. Finally, Section 6 provides

the concluding remarks.

Deep Edge Intelligence 3

2 Related Works

A signiﬁcant number of strategies have been proposed in the recent past that

focused on adapting EI into many application domains. These include video

analytics, industrial IoT, cognitive assistance, smart homes, precision agriculture

and trust prediction [2][4]. Out of the aforementioned strategies, some focused

predominantly on using DL approaches (e.g. Convolutional and Recurrent Neural

Networks, Deep Reinforcement Learning), and other traditional optimisation

techniques (e.g. Support Vector Machines), in federated and decentralised edge-

native settings [5][6][7]. Despite the increasing popularity and adaptation, almost

all these strategies paid a lack of attention to the end-to-end aspect of running

an AI strategy within a highly distributed edge environment. For instance, in

such a distributed environment, manual or even semi-automatic data labelling,

processing, model selection are often infeasible.

A comprehensive proposal for running edge intelligence strategies within a

6G environment was proposed in [8]. This work evaluates the requirements of

edge AI applications, and proposes a self-learning architecture that aims to re-

duce the degree of human intervention across multiple aspects. These aspects

include data labelling and processing, model search and consultation as well as

model retraining or tuning in the face of non-stationary data. This aligns cohe-

sively with part of our vision that aims at a more autonomous approach for edge

AI in edge environments. However, it leaves out some of the most inﬂuential

aspects of eﬃciently running an end-to-end AI strategy within an edge comput-

ing environment. These aspects include the need for self-knowledge distillation,

self-organisation of knowledge sharing topologies to share knowledge amongst

the participants of edge learning, self-healing in the face of failures as well as

automatic hyperparameter tuning.

Meanwhile, [9] proposed an EI strategy that aims to jointly optimise multi-

ple parameters to reduce the overall energy consumption of an edge IoT system.

However, the aforementioned strategy assumes that the data processing as well

as model training takes place predominantly at edge servers that are deployed

in close proximity to edge devices. Therefore, it fails to cater to scenarios where

learning happens on-device, and also does not take into account the need of

ensuring energy eﬃciency of participants (e.g. mobile devices) in a more person-

alised or targeted manner. [10] proposed a dynamic resource allocation strategy

for a decentralised federated learning setting while [11] introduced an adaptive

strategy to adaptively partition the training of a Deep Neural Network (DNN)

between the edge and device to maximise resource utilisation. All the aforemen-

tioned approaches aimed to (albeit in three distinct directions) achieve deep inte-

gration between edge and device layers, if not, the edge-device synergy. However,

individually, they fail to address the broader requirements of resource-eﬃciently

running EI strategies.

3 Vision and Systems Architecture

Below, we formally introduce the vision of DEI.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DeepEdgeIntelligence:Architecture,KeyFeatures,EnablingTechnologiesandChallengesPrabathAbeysekara1,HaiDong1,andA.K.Qin21SchoolofComputingTechnologies,RMITUniversity,Melbourne,Australiafprabath.abeysekara,hai.dongg@rmit.edu.au2ComputerScienceandSoftwareEngineering,SwinburneUniversityofTechnology,Hawth...

展开>> 收起<<

Deep Edge Intelligence Architecture Key Features Enabling Technologies and Challenges Prabath Abeysekara1 Hai Dong1 and A. K. Qin2.pdf

共15页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep Edge Intelligence Architecture Key Features Enabling Technologies and Challenges Prabath Abeysekara1 Hai Dong1 and A. K. Qin2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: