SCIENCE CHINA

Information Sciences
. REVIEW .

Pushing AI to wireless network edge: An overview

on integrated sensing, communication, and
computation towards 6G
arXiv:2211.02574v1 [cs.IT] 4 Nov 2022

Guangxu Zhu1 , Zhonghao Lyu2 , Xiang Jiao1,3 , Peixi Liu1,3 , Mingzhe Chen4 ,
Jie Xu2* , Shuguang Cui1,2,6* & Ping Zhang5,6
1
Shenzhen Research Institute of Big Data, Shenzhen 518172, China;
2
Future Network of Intelligence Institute (FNii) and School of Science and Engineering (SSE),
The Chinese University of Hong Kong (Shenzhen), Shenzhen 518172, China;
3
State Key Laboratory of Advanced Optical Communication Systems and Networks, School of Electronics,
Peking University, Beijing 100871, China;
4
Department of Electrical and Computer Engineering and Institute for Data Science and Computing,
University of Miami, Coral Gables 33146, USA;
5
State Key Laboratory of Networking and Switching Technology,
Beijing University of Posts and Telecommunications, Beijing 100876, China;
6
Peng Cheng Laboratory, Shenzhen 518066, China

Abstract Pushing artificial intelligence (AI) from central cloud to network edge has reached board con-
sensus in both industry and academia for materializing the vision of artificial intelligence of things (AIoT) in
the sixth-generation (6G) era. This gives rise to an emerging research area known as edge intelligence, which
concerns the distillation of human-like intelligence from the huge amount of data scattered at wireless net-
work edge. In general, realizing edge intelligence corresponds to the process of sensing, communication, and
computation, which are coupled ingredients for data generation, exchanging, and processing, respectively.
However, conventional wireless networks design the sensing, communication, and computation separately in
a task-agnostic manner, which encounters difficulties in accommodating the stringent demands of ultra-low
latency, ultra-high reliability, and high capacity in emerging AI applications such as auto-driving. This thus
prompts a new design paradigm of seamless integrated sensing, communication, and computation (ISCC) in
a task-oriented manner, which comprehensively accounts for the use of the data in the downstream AI appli-
cations. In view of its growing interest, this article provides a timely overview of ISCC for edge intelligence
by introducing its basic concept, design challenges, and enabling techniques, surveying the state-of-the-art
development, and shedding light on the road ahead.
Keywords Sixth-generation (6G), edge intelligence, artificial intelligence of things (AIoT), integrated sens-
ing, communication, and computation (ISCC)
Citation Zhu G, Lyu Z, Jiao X, et al. Pushing AI to wireless network edge: An overview on integrated sensing,
communication, and computation towards 6G. Sci China Inf Sci, for review

1 Introduction
With the commercialization of the fifth-generation (5G) wireless networks, we are moving towards a new
era with everything connected. The convergence of modern information and communications technology
(ICT) technologies such as the internet of things (IoT), cloud computing, mobile edge computing (MEC),
and big data analytics is prompting a huge leap in social productivity and management efficiency. At
the same time, artificial intelligence (AI) witnesses great success in various applications and continues its
* Corresponding author (email: [email protected], [email protected])
 Zhu G, et al. Sci China Inf Sci 2

(a) Number of mobile and IoT connections [7]. (b) Global data traffic growth [8].

Figure 1 Edge AI connections and data growth.

explosive growth and penetration to all walks of life. The great success achieved by AI has driven the
ongoing convergence of the communication networks with the AI technology. Particularly, in the current
5G network, AI has been used as an add-on module to boost the network performance, while in the future
sixth-generation (6G) era, AI will be deeply integrated in the network design to achieve the so-called AI-
native network. As a consequence, the future 6G wireless communication networks will go beyond a
pure data delivery pipeline, and become a comprehensive platform integrating sensing, communication,
computation, and intelligence to deliver pervasive AI services [1–3].

1.1 Edge AI

The interplay between wireless communication networks and AI technology can be generally classified into
two categories, namely the AI-assisted communication [4] and communication-assisted AI [5]. Specifically,
AI-assisted communication, which refers to the use of AI to improve existing communication systems
(e.g., physical layer modules such as channel estimation and signal detection [6]), boosts the end-to-end
performance of communication links to achieve higher rates, lower latency, and wider connectivity. On
the other hand, communication-assisted AI, which refers to the use of communication networks to help AI
acquisition, allows for distributed AI training and inference across the entire network, as well as pervasive
AI services delivery.
The focus of the current work is on the communication-assisted AI, which has received increasing atten-
tion from academia and industry in recent years, as various AI applications, e.g., industrial Internet, smart
cities, smart health, and auto-driving mature and gain popularity. Traditional communication-assisted
AI based on cloud data processing, which requires the delivery of a large amount of data collected by the
edge devices to the cloud for AI distillation, cannot support emerging mission-critical AI applications such
as industrial control, virtual reality (VR), augment reality (AR), and auto-driving due to the following
challenges.
• Massive data and network connectivity: According to Cisco, global business volumes increased
at a rate of nearly 42% per year between 2018 and 2020. As shown in Fig. 1(a), the total number of
connected devices worldwide is expected to reach 29.3 billion by 2023, with 5.7 billion mobile connections
and 14.7 billion IoT connections [7]. According to Huawei [9], the total number of global network
connections will reach 200 billion by 2030, with wireless and passive connections accounting for roughly
half of the total. Aside from the massive number of temperature, humidity, pressure, and photoelectric
sensors in the industrial sector, the network will also include a large number of smart vehicles, robots,
and drones. The constant generation of large amounts of data and network connections increases the
demand for communication capacity and computing power.
• Data sinking: Previously, big data, such as online shopping records, social media content, and
business information, was primarily generated and stored in hyperscale data centers. However, with
the proliferation of mobile and IoT devices, this trend is now being reversed. Cisco predicts that by
 Zhu G, et al. Sci China Inf Sci 3

Cloud server

• Cloud (c) Edge inference

Wide area network

• Edge servers
…
Road …
Base
Wireless AP
Inference

station
side unit

• Edge devices
… …
Autonomous driving Smart home Industry IoT
Network edge Intelligent applications
Edge AI

(a) Centralized edge learning (b) Federated edge learning

Model training
Model
aggregation

Model
training

Figure 2 The concept of edge AI and three related scenarios.

2021, all people, machines, and things will generate nearly 85 Zettabytes of usable data at the network
edge. In sharp contrast, as shown in Fig. 1(b), global data center traffic will reach only 21 Zettabytes by
2021 [8]. In the traditional cloud computing design, the massive data sinking to the network edge need
to be transferred to a central cloud server away from the edge for analysis and processing. This would
undoubtedly result in unacceptable communication cost and delay.
• Ultra-low latency requirements: The majority of new AI services require high-demand network
connections with ultra-low latency and ultra-high reliability. For example, smart industrial Internet
requires real-time state feedback, data analysis, and highly-accurate control [1]. Similarly, VR and
AR applications request real-time aggregation, analysis, and reconstruction on three-dimensional (3D)
images for complex control feedback. In these use cases, the required closed-loop sensing-communication-
computation latency must be within around one millisecond, which poses great challenges to the current
communication networks.
To address these issues, industry and academia have reached consensus on bringing computing power
and AI functionality close to data, known as edge computing or fog computing [10]. As its name implies,
edge computing aims to relocate some of the data processing and data storage for specific services from
the central cloud to the distributed edge network nodes, which are more physically and logically close to
the data provider, so as to achieve the desired low latency. AI, on the other hand, seeks to emulate human
intelligence in a machine by learning from the data. Naturally, the convergence of edge computing and
AI gives rise to a new area known as edge AI [8], which aims to provide mobile terminals with low-latency
AI services by exploiting both the computing resources and data scattered at the network edge. In view
of its promising performance gain, edge AI has received increasing attention from both academia and
industry, which is becoming a hot research area in the direction of communication-assisted AI [2, 11, 12].
As shown in Figure 2, learning in edge AI can be divided into two types, i.e., centralized edge learning
and distributed edge learning, depending on where the data is processed. After the AI model is well-
 Zhu G, et al. Sci China Inf Sci 4

Training On-device Wireless communication On-server

data computation computation

Training
Well-trained
model
Inference Real-time On-device Wireless communication On-server
data computation computation “Man
Running”
Inference result
Sensing Computation Communication Computation

ISAC Communication and

computation resource
management
AirComp

Spectrum Energy

CPU GPU Time

Figure 3 Sensing, communication, and computation in edge AI.

trained, it can be deployed for edge inference. In the early stage of the development of edge AI, AI
model is attained via centralized edge learning in many large AI companies (e.g., Google, Facebook, and
Microsoft) [11]. However, centralized edge learning requires uploading the private raw data (e.g., personal
photos at the smart phones) from edge devices to edge server, which may pose grand challenge to the
user data privacy. On the other hand, thanks to the well-known Moore’s law, the power of computing
chips such as central processing unit (CPU) and graphics processing unit (GPU) has been continuously
upgraded with decreasing cost. In particular, with the emergence of dedicated AI chipsets, the computing
power of edge devices become more and more powerful and can support the running of machine learning
(ML) tasks. This thus drives the rapid development of distributed edge learning, such as federated edge
learning (FEEL), to exploit the rich distributed computing resources at the network edge. Moreover, an
additional bonus of the FEEL is the data privacy preservation due to the waiving of the need to upload
the raw data, but sharing the less privacy-sensitive gradient or model updates.

1.2 Integrated Sensing, Communication, and Computation (ISCC)

In practice, a complex system (e.g., the edge AI system) generally consists of three coupled processes,
namely sensing, communication, and computation, as shown in Figure 3. However, in traditional wire-
less networks, these three processes are designed separately for different goals: sensing for obtaining
high-quality environmental data, communication for data delivery, and computation for executing the
downstream task within a certain deadline. Such a separation design principle encounters difficulty in
accommodating the stringent demands of ultra-low latency, ulta-high reliability, and high capacity in
emerging 6G applications such as auto-driving. This thus prompts a new wireless design paradigm of
integrated sensing, communication, and computation (ISCC) in a task-oriented manner, which compre-
hensively accounts for the use of the data in the downstream tasks (e.g., AI applications) in 6G. In the
literature, there have been some prior studies on the integration of two of the above three entities. Some
examples include joint communication and computation resource management, over-the-air computation
(AirComp), and integrated sensing and communication (ISAC), which are detailed in the following.
 Zhu G, et al. Sci China Inf Sci 5

1.2.1 Joint Communication and Computation Resource Management

Data acquisition and model computation are typically separate processes in traditional complex commu-
nication systems. With the rapid increase in data volumes in MEC, the communication and computation
capabilities at the network edge become the bottleneck. Particularly, the limited wireless resources make
it challenging for the edge server to receive significant amounts of data from edge devices swiftly through
wireless links. Hence, many researches have focused on joint communication and computation resource
management to tackle this issue in MEC. For example, in [13], in order to minimize the energy and
delay cost of the multi-user multi-task MEC system, the authors used a separable semidefinite relaxation
method to jointly optimize the offloading decision and communication resource allocation. In [14], in
order to solve the resource allocation problem for MEC, the authors proposed an effective solution to
maximize the quality of service (QoS) of all mobile devices, by transforming the problem into a linear
programming model. Besides, in [15], by using an online algorithm based on Lyapunov optimization,
the energy-latency trade-off problem of multi-user MEC systems is studied, where the computation tasks
arrive at the mobile devices in a stochastic manner. In addition, in [16], the authors solved the problem
of reducing the total energy consumption at the edge server in wireless-powered multi-user MEC systems,
by jointly optimizing the AP’s energy transmission beamforming, the user’s central processing unit fre-
quency and number of offloaded bits, and the time allocation among users to improve MEC performance.
Furthermore, joint communication and computation cooperation was exploited in MEC in [17], where
a neighboring user node was enabled as not only a relay node for helping task offloading of a user, but
also a computation helper to help remotely execute some of tasks of that user. Joint communication
and computation resource management can theoretically ensure the information security, mobile energy
saving, and so on.

1.2.2 AirComp
AirComp has emerged as a promising technology in recent years. AirComp, as opposed to “commu-
nication before computing”, integrates computing into communication, resulting in a new scheme of
“communication while computing”. In contrast to traditional wireless communication over a multi-access
channels (MAC), which requires separate transmission and decoding of information, AirComp allows
edge devices to simultaneously transmit their respective signals on the same frequency band with proper
processing, such that the functional computation of the distributed data is accomplished directly over
the air. This thus significantly improves the communication and computing efficiency, and considerably
reduces the latency required for multiple access and data fusion.
In [18], the authors provide an comprehensive overview of AirComp by introducing the basic principles,
discussing the advanced techniques and several promising applications, and identifying promising research
opportunities. In order to achieve reliable AirComp in practice, the authors in [19] focused on the power
control problem in AirComp, and the optimal power allocation under both deterministic and fading
channels were derived by minimizing the mean-squared error (MSE) of the aggregated signals. Similarly,
the authors of [20] minimized the computation MSE at the receiver by optimizing the transmitting and
receiving policy under the maximum power constraint of each sensor. While only single cell was considered
in [19] and [20], the power control problem of AirComp in the multi-cell scenario was considered in [21].
To quantify the fundamental AirComp performance trade-off among different cells, in [21], the Pareto
boundary of the multi-cell MSE region was characterized by minimizing the MSE subject to a set of
constraints on individual MSE. Note that the work in [19–21] only considered the scenario with single
antenna. The authors of [22] intended to develop multiple-input-multiple-output (MIMO) AirComp
in order to achieve multi-modal sensing with high mobility, and design MIMO-AirComp equalization
and channel feedback techniques for spatially multiplexing multi-function computation. Besides, some
complicated systems with AirComp have been also considered in the literatures. For example, the authors
of [23] considered to use reconfigurable intelligent surface (RIS) to assist AirComp. Besides, in [24], under
imperfect channel state information, the authors investigated the joint optimization of transceiver and
RIS phase design for an AirComp system. In [25], when the ground receiver is unavailable, unmanned
 Zhu G, et al. Sci China Inf Sci 6

aerial vehicles (UAVs) are utilized to establish line-of-sight (LoS) connections by tracking mobile sensors,
and thus improving the performance of AirComp.

1.2.3 ISAC
ISAC generally refers to the integration of sensing and communication into a unified design in wireless
networks to enhance the utilization efficiency of scarce spectrum and wireless infrastructures, pursue a
mutual benefit via sensing-assisted communication and communication-assisted sensing [26]. In compari-
son to traditional wireless networks, ISAC can use the wireless infrastructure as well as limited spectrum
and power resources for both communication and sensing, which could improve the system performance
at a lower cost.
ISAC is one of the potential key technologies in 6G networks that has received a lot of attention in
the literature. Many works have focused on joint sensing and communication in [27]. For example,
in [28], the authors proposed a dual-functional MIMO radar communication system that consists of a
transmitter with multiple antennas that can communicate with downlink cellular users and detect radar
targets at the same time. In [29], a joint transmit beamforming model for a dual-function MIMO radar
and multiuser MIMO communication transmitter was proposed and the weighting coefficients of the radar
beamforming were designed. In [30], the authors also considers the beamforming optimization problem
in ISAC system, and the radar sensing performance is maximized subject to the communication users’
minimum signal-to-interference-plus-noise ratio (SINR) requirements and the transmit power constraint
of the base station (BS). The authors in [31] employ the Cramer-Rao bound (CRB) as a performance
metric of target estimation, and the CRB of radar sensing is minimized while guaranteeing a pre-defined
level of SINR for each communication user. In [32], the authors considered a UAV-enabled ISAC system,
where UAV trajectory/deployment and beamforming design are jointly considered to balance the sensing-
communication performance trade-off under quasi-stationary and mobile UAV scenarios, respectively.
Furthermore, because of the benefits of spectrum sharing, ISAC has been used in a variety of systems.
Furthermore, using RIS to facilitate radar sensing and ISAC has attracted growing research interests
(see, e.g., [33–37]). For instance, the authors in [33] derived the fundamental CRB for RIS-enabled
NLOS sensing, and those in [34] jointly designed the transmit beamforming at the transmitter and the
reflective beamforming at the RIS for ensuring both sensing and communication performances. The
authors of [35] used RIS for a joint design of constant-modulus waveform and discrete phase shift to
mitigate multi-user interference in ISAC. In addition, in the overview paper [36], the authors elaborated
the benefits of RIS in wireless communication, sensing, and security, and envisioned that the RIS-assisted
communication and sensing would mutually benefit each other. The authors of [37] considered the
combination of ISAC and AirComp to improve the spectral efficiency and sensing performance, and
the beamformers for sensing, communication, and computation were jointly optimized. The authors
of [38] used ISAC in smart homes to provide inconspicuous sensing and ubiquitous connectivity. Besides
joint sensing and communication, there are also a line of research on sensing-assisted communication. For
example, a radar-assisted predictive beamforming design for vehicle-to-infrastructure communication was
investigated in [39], and it was found that the communication beam tracking overheads can be drastically
reduced by exploiting the radar functionality of the road side unit.

1.2.4 Task-oriented ISCC Towards Edge AI

As previously stated, sensing, computation, and communication have a symbiotic relationship, especially
in the context of edge AI. Specifically, the ultimate performance of edge AI depends on the input feature
vector’s distortion level arising from three processes, i.e., data acquisition (sensing), feature extraction
(computation), and feature uploading to edge server (communication). Particularly, sensing and com-
munication compete for radio resources, and the allowed communication resource further determines the
required quantization (distortion) level such that the quantized features can be transmitted reliably to
the edge server under a delay constraint. Thereby the three processes are highly coupled and need to
be jointly considered. Furthermore, the implementation of ISCC should be designed under a new task-
 Zhu G, et al. Sci China Inf Sci 7

Edge AI

Sec. I: Introduction Sensing, Communication and Computation Integration

Joint Communication and Computation Resource

Management

Mixup Data Augmentation with AirComp

Sec. II: Centralized Edge Learning Centralized Learning with ISAC

Research Opportunities

Joint Communication and Computation Resource

Management

Air-FEEL

Structure of the Survey Sec. III: Federated Edge Learning Federated Edge Learning with ISAC

Research Opportunities

JSCC in Edge Inference

Joint Communication and Computation Resource

Management

Sec. IV: Edge Inference Over-the-air Edge Inference

Co-inference with ISAC

Research Opportunities

Sec. V: Concluding Remarks

Figure 4 The structure of this article.

oriented principle that concerns the successful completion of the subsequent AI task. Different from
conventional communication system design aiming at maximizing the data-rate throughput, the ultimate
performance metrics of interest for the system become the inference/training accuracy, latency, and en-
ergy efficiency, etc. For instance, an edge AI task-oriented ISCC scheme can be designed to maximize the
inference/training accuracy under constraints on low latency and on-device resources. This is in sharp
contrast to the classic separation-based design approach that consider the sensing, communication and
computation processes in isolation. More examples of the task-oriented ISCC will be discussed in the
subsequent sections.

1.3 Structure of the Survey

Different from prior works considering the integration of sensing and communication or computation and
communication in generic wireless networks, we focus on their integration towards edge AI applications.
As a result, we classify them based on three application scenarios, i.e., centralized edge learning, federated
edge learning, and edge inference. The remaining of the survey is organized as shown in Figure 4. In
Section 2, we first review the joint communication and computation resource management, mixup data
augmentation with AirComp, and ISAC, respectively, in centralized edge learning, and then discuss
the related research opportunities. In Section 3, we discuss the joint communication and computation
resource management in federated learning, the application of AirComp into FEEL, and the combination
of ISAC and FEEL, as well as the related research opportunities. In Section 4, we first discuss the joint
source and channel coding, then present joint communication and computation resource management in
edge inference, over-the-air edge inference, and co-inference with ISAC, respectively, and highlight several
key research opportunities. Finally, Section 5 concludes the article.
 Zhu G, et al. Sci China Inf Sci 8

2 Centralized Edge Learning

With the continuing development of deep learning (DL) in recent years, the increasing DL model com-
plexity has posed a grand challenge in training due to the demand for computation power and storage
capacity. There are two basic strategies to accommodate the increasing demands for the resources re-
quired by DL at the centralized cloud: scaling-up, which involves adding extra processing and storage
resources to a single central server, and scaling-out [40], which involves forming a server cluster by net-
working multiple servers each with certain computing and storage capacity. Recent rapid development
of MEC makes it possible to deploy the mentioned two strategies at the network edge so as to “bring the
computation power close to the data”, leading to an emerging research area known as centralized edge
learning.
In the centralized edge learning, the data is first collected on the client side and then transferred through
the wireless channel to a central edge server. The central server then stores and processes the data, and
finally returns the learnt model back to the client. This architecture is simple to deploy and manage,
particularly in circumstances where data is scattered across geographically disperse nodes. However, due
to the need for centralized data processing, long delays and high transmission costs accompany when
the communication channel capacity between the client and the central server is limited. Furthermore,
due to the central edge server’s limited computational power and storage resources, it is difficult to
enable the construction of complicated models based on massive datasets using centralized edge learning.
Therefore, in order to improve the efficiency of centralized edge learning, the sensing, communication,
and computation processes need to be jointly designed and the associated resources should be judiciously
managed as described in the sequel.

2.1 Joint Communication and Computation Resource Management in Centralized Edge

Learning

As previously stated, centralized edge learning may introduce a significant delay, which will be catas-
trophic in latency-sensitive applications such as autonomous driving and VR games. Furthermore, the
massive data communication places a significant strain on the backbone network, resulting in significant
computation overhead for the central server. Thus, centralized edge learning exists at the crossroads
of two domains: communication and computing. The advent of new disciplines brings many interdisci-
plinary research opportunities, and joint design is required to interweave the two sectors in order to solve
the aforementioned challenges to accomplish fast and efficient intelligence acquisition while allocating
suitable resources. The overall goal of traditional communication systems is to maximize throughput.
Differently, centralized edge learning systems aim to maximize learning performance. As a result, if the
traditional communication system optimization method is still used, optimal learning performance may
not be achieved because specific learning factors such as model and data complexity are not taken into
account. For example, because the structures of a support vector machine (SVM) and a convolutional
neural network (CNN) differ, they can achieve different learning accuracies with the same training data
sample size. Furthermore, the communication cost of transferring a data sample in various ML tasks
can vary significantly. As a result, new resource allocation algorithms are required for centralized edge
learning.
There have been several prior works proposing optimized resource allocation schemes for centralized
edge learning in recent years. One efficient design is to schedule the joint communication and computation
processes based on the data importance. In [41], the authors proposed a data-importance-aware user
scheduling scheme for edge ML systems by considering specific SVM tasks. In this scheme, a data
importance measure is proposed to classify the data into different importance levels and then allocate
more resources to the data with higher importance. Furthermore, the design based on SVM in [41] was
then extended to CNN, in which the authors considered the specific retransmission decision problem
in order to ensure the quantity and quality of training data in the face of transmission data errors.
Unlike the traditional automatic-repeat-request which focuses solely on reliability, the authors selectively
retransmitted data samples based on their importance in order to accelerate the convergence speed
 Zhu G, et al. Sci China Inf Sci 9

Figure 5 Transmitting procedure in AirMixML.

and accuracy of the training. The other line of research focuses on the learning-centric wireless resource
allocation. In [42], the authors proposed and validated a nonlinear classification error model for ML tasks.
A power allocation scheme centered on learning performance was proposed based on this model. Using
the total variational minimum optimization method, the Karush-Kuhn-Tucker (KKT) solution could be
obtained. Furthermore, in [43], the authors employed statistical mechanics of learning to forecast the
relationship between the learning accuracy of various tasks and the amount of training data. The learning
performance is maximized by using differential convex programming (DCP). According to the analysis,
the optimal transmission time is inversely proportional to the generalization error.

2.2 Mixup Data Augmentation with AirComp

AirComp is another integrated communication-and-computation design that can be exploited to enhance

the efficiency of centralized edge learning, and particularly combined with the mixup data augmentation.
Mixup is a well-known data enhancement technology [44]. Mixup, which in essence, trains neural networks
using convex combinations of data and labels, thus regularizing neural networks to make learning simple
linear behaviors between training samples easier.
As shown in Fig. 5, the authors in [45] combined mixup and AirComp, in which a ML framework
called over-the-air mixup ML (AirMixML) is proposed to take advantage of the natural distortions and
superpositions properties of wireless channels. Multiple users in AirMixML send simulated modulated
signals of their private data samples to a central server, which uses the received aggregated noisy samples
to train ML models, while protecting the user’s privacy. AirMixML was shown to train the model
using mixup data enhancement and achieves the same accuracy as the raw data samples. Specifically,
AirMixML adjusts the privacy disclosure level of the transmitted signal by controlling the user’s transmit
power, and the Dirichlet dispersion ratio controls the local power contribution of each worker to the
superimposed signal. Given that AirMixML is the first privacy-preserving edge ML framework to use
over-the-air signal superposition and additional channel noise without on-device training, there is still
much room for improvement before practical application.

2.3 Centralized Edge Learning with ISAC

The traditional sensing and communication stages are carried out sequentially. Nevertheless, in ISAC
systems, the sensing and communication are implemented via unified wireless signals. This thus motivates
many prior works that proposed to accelerate the centralized edge learning by ISAC. In particular, the
sensing and communication stages are combined to maximize the use of wireless signals for both dataset
generation and transmission. However, ISAC also introduces additional interference between sensing and
communication functions to be dealt with.
In order to integrate dataset generation and uploading, in [46], the authors proposed a classification
error minimization method for beamforming and time allocation. Besides, several prior works combined
ISAC with data offloading in MEC, which can be utilized in the centralized edge learning with ISAC.
 Zhu G, et al. Sci China Inf Sci 10

In [47], the user conducted sensing and communication on the same spectrum using a MIMO array
based on the dual-functional radar communication technology. The authors proposed a multi-objective
optimization problem for jointly optimizing transmit precoding for sensing, communication and allocating
computation resources, by considering both beampattern design and multi-user MIMO radar computation
offloading energy consumption. In [48], the authors studied the throughput maximization in a multi-user
MEC system using a sense-then-offload protocol. Furthermore, in [49], the authors investigated a traffic-
aware task offloading scheme in a vehicular network and proposed an offloading mechanism based on the
sensed environment data. In [50], the author proposed a sensory system for sensing and communication
based on analog spike signal processing.

2.4 Research Opportunities

Despite the research efforts discussed above for efficient centralized edge learning, there are still many
unexplored territories yet, which are discussed as follows.
• Secure Data Uploading: The centralized edge learning architecture may suffer from the single
point of failure issue, that is, the central server is prone to attack by malicious users who may upload
forged or poisoned data to mislead the entire training process. Therefore, how to build a trustworthy
mechanism to guard against the potential attack in the data uploading process is a critical issue warrant
further study.
• Data-importance aware systems: Due to the limited communication resources in a real com-
munication system, it is often not possible to transmit all of the datasets, so it is necessary to selectively
transmit the data in order to train the network more efficiently. Furthermore, the importance of data
changes during the transmission process (for example, a certain large category of data is important until
it is transmitted, but after a certain amount of data has been transmitted, the same large category of
data has little impact on the training), and this aspect has not been considered in previous studies.
• Task-oriented ISCC in centralized edge learning: The various learning tasks in centralized
edge learning frequently require the simultaneous support of sensing, communication, and computation
functions. As a result, joint resource management for ISCC is required to improve the performance of
centralized edge learning. There is still an unexplored territory for task-oriented ISCC targeting edge
inference, which motivates the future work’s main theme.

3 Federated Edge Learning

FEEL is a machine learning setting where multiple edge devices collaborate in training a shared ML
model, and each device’s raw data is stored locally and not exchanged or transferred. From the networking
perspective, FEEL can be divided into two classes, including centralized FEEL and decentralized FELL.
Centralized FEEL is the most popular FEEL architecture, in which an edge server coordinates the ML
model training among the edge devices. Unlike centralized FEEL, decentralized FELL is a network
topology without any central server to coordinate the training process, in which all edge devices are
connected in a peer-to-peer manner to perform the model training. For FEEL, the sensing, communication
and computation are three important functions that must be involved in completing a learning task using
FEEL. To begin with, the edge devices must sense the environment (e.g., by using the equipped radio
sensors) in order to obtain data for training ML models; Second, the edge devices compute model updates
using local computing power; Finally, the edge devices upload local model updates to the edge server
via the uplink channel, and the edge server broadcasts the global model to each edge device via the
downlink channel. Wireless communication, sensing, and computation all have different effects on federal
edge learning, however, in edge networks where resources (e.g., radio bandwidth and edge device energy
supply) are typically limited, and the three are frequently coupled and constrained by each other, thus
making their joint design necessary.
 Zhu G, et al. Sci China Inf Sci 11

3.1 Joint Communication and Computation Resource Management in FEEL

In FEEL, the heterogeneity of edge devices in terms of radio sources, channel status, and computational
capabilities is a common issue that has a direct impact on the ultimate learning performance such as
accuracy and latency. This thus calls for joint communication and computation resource allocation for
FEEL performance optimization. Specifically, the mentioned device heterogeneity essentially leads to
distinct uploading time between different devices and edge server. As a result, if the edge server uses
synchronized aggregation of the model updates of the edge devices, the edge device with the longest
delay dominates the communication time of a single round. To address this issue, the authors in [51]
investigates the optimal scheduling scheme for edge devices to minimize the training time of federal
edge learning, but the communication resource optimization is not involved. In [52], the heterogeneous
channel conditions of edge devices are also considered, in which the objective is to maximize the number
of participating edge devices, by optimizing the scheduling scheme of edge devices and the allocation
of communication resources including transmit power and bandwidth. In addition to the heterogeneity
in channel conditions, the edge devices tend to be heterogeneous in computing capabilities as well. For
instance, both [53] and [54] comprehensively consider the heterogeneity of edge devices in terms of channel
conditions and computing capabilities, and study the optimal scheduling scheme for edge devices and
the optimal communication resource allocation. The difference between the [53] and [54] is that the
considered problem in [53] only focuses on a single communication round and each communication round
is treated equally. In contrast, [54] explicitly takes the importance of different communication rounds
into consideration and investigates the bandwidth allocation and edge device scheduling problems under
long-term energy constraints. Moreover, imperfect wireless channel conditions are also investigated in [54].
On the other hand, as edge devices need to utilize local computing resources for model updates. Re-
search on computing resource management mainly focus on two directions: optimization of the CPU/GPU
frequency of edge devices [55,56] and optimization of the batch size used in model updates (e.g., stochastic
gradient descent) [57,58]. Since edge devices are usually heterogeneous in terms of computing power, the
time for completing training and the energy consumed by federal edge learning can be largely wasteful if
the computing power of different edge devices is not reasonably tuned. Both the literature [55] and [56]
consider the total energy consumption of the system and the required training time simultaneously as
minimization objectives to optimize the CPU frequency of different edge devices, as well as system vari-
ables such as communication resources and device scheduling. For the case where the edge devices cannot
effectively adjust the computing frequency, [57] optimizes the batch size of different devices to align the
communication delay between different edge devices and the edge server, thus reducing the training time
for federal edge learning. Unlike [57], which only considers the optimization within a single communi-
cation round, literature [58] focuses on the whole federated edge learning training process, where the
authors consider a long-time dynamic resource optimization problem, and a scheme based on Lyapunov
optimization is proposed to jointly optimize the computing frequency and the batch size of each edge
device in different communication rounds.
Besides the heterogeneity of communication and computation, the data heterogeneity is also typical in
federated learning, as the data at each edge device is likely to be personalized towards the specific device.
A few works tended to manage the communication and computation resources in FEEL, by considering
the effects of data heterogeneity. For example, in [59], the optimal client sampling strategy that tackles
both system and data heterogeneity is designed to minimize the training time with convergence guarantees
in a FEEL system with resource-constrained devices. The work in [60] considered quantized FEEL with
data heterogeneity, and jointly optimized the quantization level and the bandwidth allocation to minimize
the training time.

3.2 Over-the-Air Federated Edge Learning (Air-FEEL)

Air-FEEL has emerged as a promising solution for edge AI [61]. As shown in Figure 6, over-the-air
model/gradient aggregation is used in Air-FEEL to improve learning efficiency of FEEL. It is shown in [62]
that, compared with the conventional orthogonal multi-access, Air-FEEL can reduce the communication
 Zhu G, et al. Sci China Inf Sci 12

latency by a factor approximately equal to the number of devices without significant loss of the learning
accuracy. Various research efforts have focused on different directions in Air-FEEL, such as resource
management, device scheduling, and privacy preserving schemes.
In Air-FEEL, edge devices can control their transmit power adaptively to reduce aggregation error
for model/gradient aggregation, and thus improve learning accuracy or convergence rate. The authors
in [63] developed the optimized power control to minimize the AirComp MSE, which has a threshold-
based structure depending on the channel conditions at different edge devices. Rather than minimizing
the aggregation MSE in each round independently, the authors in [19, 64] optimized power allocation
across multiple global rounds to maximize the convergence rate in terms of the loss optimality gap.
Device scheduling is another efficient technique for improving Air-FEEL performance via addressing
resource heterogeneity by dropping edge devices with poor communication and computation conditions.
The work [65] presented a joint design of device scheduling and receive beamforming in which the edge
devices with weak signal strengths after receive beamforming were dropped from the training process.
Furthermore, the work [66] investigated device scheduling by taking into account their diverse energy
constraints and computation capabilities, and an energy-aware dynamic device scheduling algorithm
based on Lyapunov optimization was proposed. Utilizing Air-FEEL offers an additional advantage in
improving data privacy, in addition to the benefit of reducing multiple-access latency. The local training
data can still be inferred from the local model updates even with powerful model inversion techniques [67].
As a fix, Air-FEEL limits eavesdroppers’ access to the aggregated updates, hiding each individual local
update in the sea of others. A further free mask that can be used to safeguard the privacy of the
data is the random disturbance that the channel noise imposes on the aggregated updates [68]. A more
comprehensive overview on Air-FEEL can be referred to [61].
Although AirComp is beneficial for model aggregation in Air-FEEL due to the inherent superposition
property of wireless channels, Air-FEEL also suffers from the straggler issue in which the device with
the weakest channel acts as a bottleneck of the model aggregation performance. To address this issue,
the work in [69] leveraged the RIS technology in Air-FEEL, and a unified communication-learning opti-
mization problem is solved to jointly optimize device selection, over-the-air transceiver design, and RIS
configuration. The aforementioned studies all focused on the centralized FEEL, in which a central edge
server is required to orchestrate the training process. The work in [70,71] considered decentralized FEEL
in the scenario where an edge server is not available or reliable, in which the authors considered the
precoding and decoding strategies for device-to-device communication-enabled model/gradient aggrega-
tion and proposed an AirComp-based decentralized stochastic gradient decent with gradient tracking and
variance reduction algorithm to reach the consensus.

3.3 Federated Edge Learning with ISAC

Federated edge learning also be exploited to training AI models in wireless sensing networks, and in this
case, ISAC can be integrated to enhance the training and sensing performance, as shown in Figure 7.
However, it is still an open problem to design edge learning systems with sensing and communication
coexistence [26].
In ISAC, sensing and communication can be integrated in three different levels. At the first level, the
sensing and communication signals may occupy orthogonal time-frequency resources, which do not inter-
fere with each other functionally but compete for time-frequency resources, so the reasonable resource
allocation between sensing and communication is a key issue when communication and sensing orthog-
onally coexist. For example, in [72], the sensing and communication work in a time-division manner,
so that sensing and communication jointly compete for time resources. Based on this, a mathematical
relationship between sensing time and learning performance is experimentally fitted and the trade-off
between learning performance and communication rate is analyzed in [72]. The work [73] considered hor-
izontal FEEL with ISAC and jointly optimizes the sensing, communication and computation resources,
where sensing and communication occupy the same frequency band and different time durations, similarly
in [72]. At the second integration level, sensing and communication may work on the same time-frequency
 Zhu G, et al. Sci China Inf Sci 13

Uplink transmission Edger device

Downlink broadcast

Edger server

Analog
aggregation

Figure 6 Illustration of the Air-FEEL system.

Downlink broadcast
ISAC signal
Edger device Sensing target

Edger server

Model
aggregation

Figure 7 Illustration of the FEEL system with ISAC.

resources but use separated signal waveforms, and they will interfere with each other, so how to manage
interference becomes especially important. In [74], the relationship between learning performance and
sensing/communication resources is obtained by considering sensing and communication working on the
same time-frequency resources and linking the learning performance to the quantity of sensed samples.
Since sensing and communication may interfere with each other spatially, in [74], the beam directions
of the perception and communication signals are optimized with the goal of maximizing the learning
performance. In the third level, sensing and communication functions are simultaneously implemented
using a shared signal waveform, which can effectively improve the system spectrum efficiency, hardware
efficiency, and information processing efficiency [26, 75, 76]. The literature [77] is the first to combine
level-three ISAC with over-the-air analog aggregation technology and jointly designed beamforming for
both sensing and communication signals, laying the foundation for the subsequent application of ISAC
to federated edge learning. Unleashing the full potential of FEEL by combining with ISAC has very
significant application prospects. In order to save frequency resources, the same ISAC signal is adopted
for both sensing and communication in [78], and a cooperative sensing scheme based on vertical FEEL
is proposed to enhance the sensing performance.
 Zhu G, et al. Sci China Inf Sci 14

3.4 Research Opportunities

Despite the research efforts discussed above for efficient FEEL, here lists some unexplored problems and
challenges to motivate future works.
• Multi-modal data sources: In real-world sensing systems, the edge devices involved in training
may have different types of sensors, such as radio sensors or cameras, and the sensed data may have
different modalities [79]. In such cases, a FEEL system with ISCC for multimodal data needs to be
designed and optimized.
• Dynamic sensing environments: Most of the current works consider static sensing environments,
but in real scenarios the sensing environment may keep changing over time and the distribution of the
sensed data samples will no longer be stationary. How to design a federated edge learning ISCC system
for dynamic sensing environments is also a topic worthy of in-depth study.
• Task-oriented ISCC in FEEL: Few current studies on FEEL have considered specific sensing
process, mostly focusing on resource optimization for communication and computation. Various types of
learning tasks in FEEL often require the support of sensing, communication, and computation functions
at the same time, resulting in a variety of complex relationships among the sensing, communication, and
computation modules, such as coupling, collaboration, and even competition. Therefore, in order to im-
prove the upper limit of the performance of FEEL, joint resource management of sensing, communication,
and computation is required.

4 Edge Inference

Apart from the training phase discussed above, edge inference is another important aspect for supporting
the successful implementation of AI technologies at wireless edge [80]. Specifically, for edge inference,
a well-trained ML model needs to be deployed at network edge to run AI tasks in real time (such
as classification, recommendation, and regression), which is beneficial to computation/storage/power-
limited edge devices and delay-sensitive AI tasks [81]. Thus, edge inference has become an important
technique to enable various AI applications, such as auto-driving and smart cities in 6G networks. For
example, in auto-driving, the vehicles need to detect obstacles to avoid accidents under stringent latency
constraints. To guarantee high detection accuracy, more sophisticated DNN architectures are preferable
for the detection, say, ResNet-50 [82]. However, ResNet-50 contains 50 convolutional layers, and demands
nearly 100 megabytes of memory for storage. On one hand, it is non-trivial to deploy such complicated
DNN on edge devices with limited computation and storage capacity. In this case, it merely on the cloud
server as this will induce intolerable delay and may increase the risk of accidents. To deal with such
dilemma, edge inference provides a promising solution with better trade-off among computation power,
storage, and communication latency.
There are three different methods to implement edge inference, namely on-device inference [83], on-
server inference [84, 85], and split inference [86–99]. For on-device inference, the computation is ac-
complished merely on edge devices, which is non-trivial for recent increasingly complex AI models and
computation/storage/power-limited edge devices. To tackle such issue, on-server inference is designed.
However, on-server inference suffers from the communication bottleneck due to the potential high-volume
data transmission over the band-limited wireless channels in the presence of uncertain channel fading,
under the stringent low-latency constraint. Also, the computation resources at edge servers may still
fall short when running some large-scale AI tasks. Nevertheless, potential information leakage during
data uploading from the edge devices to the edge server may lead to privacy issues in edge inference.
To tackle these problems, split inference is proposed jointly considering techniques such as joint source
and channel coding (JSCC) [86–91, 100], joint communication and computation resource management
design [92–96], and AirComp [97, 98]. Furthermore, as a recently proposed technique, ISAC has drawn
increasing research interests [26]. Intuitively, ISAC can further reduce the latency of edge inference due
to the integrated data sensing and uploading process [99]. The joint management of sensing, communi-
 Zhu G, et al. Sci China Inf Sci 15

cation, and computation resources in this case is more difficult. In the following, we discuss the above
techniques in detail.

4.1 JSCC in Edge Inference

Generally, when uploading features from the devices to the serves to perform inference, the fluctuating
wireless channels may introduce lossy transmission, which calls for proper design of JSCC for efficient
feature transmission. Moreover, in some sense, JSCC can be viewed as a novel design principle bringing
both communication and computation into consideration. With the recent development of DL, deep JSCC
has been widely investigated to alleviate excessive signaling overhead as well as improve the robustness
to channel distortion. For example, for classification tasks, the authors in [86] and [87] proposed a
retrieval-oriented wireless image transmission framework to maximize the classification accuracy, where
the JSCC framework is trained by the cross-entropy between the predictions and the ground-truth labels.
Moreover, information bottleneck (IB) [100] was proposed to extract minimum features to fulfill certain
tasks sufficiently. Under the guidance of IB, an image classification task was considered in [88] by designing
a framework of task-oriented JSCC. By combining IB with stochastic optimization, the same task was
considered in [89] to minimize energy consumption as well as service latency simultaneously. Considering
edge inference with multiple edge devices, a task-oriented JSCC framework was designed in [90], where
a group of edge nodes perform the classification task coordinated by an edge server. Furthermore, some
initial exploration has been made in [91] to deal with the multi-modal data, where a task-oriented semantic
communication scheme was proposed and the cross-entropy objective with multiuser multi-modal data
fusion was considered.

4.2 Joint Communication and Computation Resource Management in Edge Inference

Edge inference typically involves local feature extraction and uploading to the edge server for further
processing, which may encounter both communication and computation bottleneck. Specifically, on one
hand, transmitting data through wireless links naturally suffers from channel impairment, especially
when the AI services have stringent low-latency requirement. On the other hand, processing data at
devices and servers incurs computational delay, especially when a large number of devices require for AI
services simultaneously. Furthermore, there always exists a trade-off between inference accuracy and the
computation-communication capability of the system in edge inference.
To address the above issues, joint management of communication and computation resources is con-
sidered in recent works. For instance, in [92], the optimal control of inference accuracy and transmission
cost was modeled as a Markov decision process (MDP), and their trade-off is balanced via dynamically
selecting the optimal compression ratio with hard deadline requirements. By further considering proper
split of AI models in multi-user edge inference systems, the authors in [93] jointly designed the model
split point selection and computational resource allocation to minimize the maximum inference latency.
Also, the authors in [94] studied the trade-off between the computational cost of the on-device model and
the communication overhead of uploading the extracted features to the edge server. Then they propose
a three-step framework for inference, which contains model split point selection, communication-aware
model compression, and task-oriented encoding mechanism for the extracted features. Due to the heavy
computational burden being offloaded on the server, it is also urgent to study the computation and in-
ference accuracy trade-off at the server side. In [95], the authors considered the early exiting technique,
which allows a task exit from certain layers of a DNN without traversing the whole network. In such way,
the joint management of communication and computation resources could reduce the inference latency
under various accuracy requirements. Moreover, a progressive feature transmission protocol was pro-
posed in [96], which contains importance-aware feature selection and transmission-termination control.
In such a protocol, the devices transmit the extracted features progressively according to their impor-
tance, and once the inference accuracy requirement is obtained, the transmission will stop. With such
design, the trade-off between inference accuracy and communication-computation latency can be well
balanced. Moreover, RIS has recently emerged as a potential solution to provide a cost-effective way for
 Zhu G, et al. Sci China Inf Sci 16

Figure 8 Illustration of AirComp-based edge inference systems.

enhancing the performance of edge inference. For example, a RIS-aided green edge inference system was
considered in [101], where the set of tasks performed by each BS, uplink/downlink beamforming vectors
of BSs, transmit power of edge devices, and uplink/downlink phase-shift matrices at the RIS were jointly
designed to minimize the overall network power consumption.

4.3 Over-the-air Edge Inference

AirComp is also appealing for low-latency edge inference by seamlessly integrating communication and
computation, as shown in Fig. 8. The research on over-the-air edge inference is still in its early stage.
An initial study of AirComp-based multi-device edge inference system was made in [97], where Air-
Comp is utilized to aggregate multiple noisy feature observations of a common source to average out the
feature noise for boosting the inference accuracy. The authors first characterize the influence of sensing
and channel noise on inference accuracy by deriving a tractable surrogate performance metric called dis-
criminant gain. Then the authors maximize the inference accuracy by jointly optimizing the transmit
precoding and receive beamforming. Besides the significantly enhanced spectrum efficiency, AirComp
has additional benefit in privacy preservation. To exploit such property, the authors in [83] considered
an ensemble inference framework, where each device needs to transmit their predictions to the server
for further fusion to obtain final results. Apart from maximizing inference accuracy, the authors also
consider maximizing the privacy of the on-device models. To this end, AirComp is exploited for privacy-
enhanced outcome fusion as each individual predictive outcome is hidden in the crowd. Specifically,
the authors introduce different ensemble methods, such as belief summation and majority voting, and
provide privacy analysis for these AirComp-based fusion schemes. Finally, numerical results provided
in [83] have shown that the proposed AirComp-based solution significantly outperforms other orthogonal
transmission schemes in terms of the required communication overhead under the same target privacy
guarantee. Moreover, the authors in [98] also considered the privacy issues in edge inference. Specifically,
they consider the distributed inference of graph neural networks (GNNs). To deal with the possible
privacy leakage problem arising from the devices exchanging information with neighbors during decen-
tralized inference, the authors first characterize the privacy performance of the considered decentralized
inference system. Then they design privacy-preserving signals and the corresponding training algorithms
in combination with AirComp to further boost the privacy of the considered system.
 Zhu G, et al. Sci China Inf Sci 17

Figure 9 Edge inference systems with multi-device sensing.

4.4 Co-inference with ISAC

In future wireless networks, to support environment-aware intelligent applications, it is desirable to

process and upload the collected data from sensing devices for inference, where sensing, communication,
and computation are naturally coupled and need to be jointly designed. However, integrating ISAC with
edge inference introduces several issues. First, it is non-trivial to characterize the inference performance
in ISAC-enabled edge inference. Second, how to jointly design the resources for sensing, communication,
and computation to proper balance the trade-off among them is challenging. To deal with the above
issues, the authors in [99] studied a task-oriented ISCC-based edge inference system as shown in Fig.
9, where multiple ISAC devices collect sensing data, and then upload the quantized features to the
server for classification. The authors analyze inference performance in such ISCC-based system via
deriving the tractable discriminant gain, based on which, the allocation of sensing, transmit power,
communication time, and quantization bits are jointly designed for the successful completion of the
subsequent classification task. Finally, some interesting design insights for balancing the trade-off between
sensing, communication and computation were crystalized in [99]: the sensing power and quantization
bits should be enlarged as the number of classes increases in the classification task, otherwise, more
communication power should be allocated if the channel conditions of the devices are poor.

4.5 Research Opportunities

Despite the research efforts discussed above for efficient edge inference, there are still many open problems
and challenges unexplored yet.
• Fundamental limits of JSCC: Similar as Shannon’s information theory, the fundamental limits of
semantic based JSCC transmission for inference needs to be characterized. Moreover, it is also interesting
to explore the number of optimized symbols for the successful completion of certain tasks via JSCC
transmission, which can balance the delay and accuracy trade-off in JSCC-based edge inference.
• Case with multiple devices and/or multiple servers: For edge inference systems with multiple
devices and/or multiple servers, the device selection and server coordination need to be further considered.
On one hand, the selection of devices for inference need to account for the trade-off between delay and
accuracy. On the other hand, multi-server system requires flexible model deployment for large-scale AI
tasks, such as heterogeneous tasks with different models requirements.
• Task-oriented ISCC for edge inference: In edge inference, sensing, communication, and compu-
tation may compete for resources (such as radio and hardware). How to depict the relationship between
the inference performance and all the three mentioned processes is quite challenging, thus yielding the
 Zhu G, et al. Sci China Inf Sci 18

non-trivial problem of joint management of sensing, communication, and computation resources. Al-
though [99] did an initial study in this direction, there still remain many uncharted issues warrant further
investigation. Moreover, to fully exploit edge intelligence, multi-modality sensors (such as laser radar,
millimeter-wave radar, and cameras) may be deployed at wireless edge. How to process the acquired
multi-modal sensing data for efficient inference is also an interesting direction to pursue.

5 Concluding Remarks

In the upcoming 6G era, we shall witness the paradigm shift in network functionality from connecting
people and things to connecting intelligence, driving the advancement of IoT in 5G towards AIoT. As
a consequence, with the assist of 6G networks, AI is expected to spread from the cloud to the network
edge so as to deliver pervasive AI services. However, classic design principle of separating sensing,
communication, and computation cannot meet the stringent demand in terms of latency, reliability, and
capacity.
To tackle this issue, this article presented a timely literature survey on the tasked-oriented ISCC
for edge intelligence. First, we introduced the motivations and basic principles on ISCC. Then, we
introduced representative works on three different scenarios, i.e., centralized edge learning, federated
edge learning, and edge inference, respectively, by focusing on the joint communication and computation
resource management, AirComp, and ISAC, in each scenario. Finally, interesting research directions
were presented to motivate future works. It is our hope that this article can provide new insights on this
interesting research topic, and motivate more interdisciplinary research from the communities of wireless
sensing, wireless communications, machine learning, and computing.

References
1 X. You, C. Wang, J. Huang, X. Gao, Z. Zhang, M. Wang, Y. Huang, C. Zhang, Y. Jiang, J. Wang, et al., “Towards
6G wireless communication networks: Vision, enabling technologies, and new paradigm shifts,” Science China
Information Sciences, vol. 64, no. 1, pp. 1–74, Nov. 2021.
2 K. B. Letaief, Y. Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6G: Vision, enabling technologies, and
applications,” IEEE J. Select. Areas Commun., vol. 40, no. 1, pp. 5–36, Nov. 2022.
3 Z. Feng, Z. Wei, X. Chen, H. Yang, Q. Zhang, and P. Zhang, “Joint communication, sensing, and computation
enabled 6g intelligent machine system,” IEEE Network, vol. 35, no. 6, pp. 34–42, Dec. 2021.
4 K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y.-J. A. Zhang, “The roadmap to 6G: AI empowered wireless
networks,” IEEE Commun. Mag., vol. 57, no. 8, pp. 84–90, Aug. 2019.
5 X. Shen, J. Gao, W. Wu, M. Li, C. Zhou, and W. Zhuang, “Holistic network virtualization and pervasive network
intelligence for 6G,” IEEE Commun. Surveys Tuts., vol. 24, no. 1, pp. 1–30, Dec. 2022.
6 H. Ye, G. Li, and B. Juang, “Power of deep learning for channel estimation and signal detection in OFDM systems,”
IEEE Wireless Commun. Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018.
7 “Cisco annual internet report (2018–2023),” Tech. Rep.
8 “Cisco global cloud index: Forecast and methodology, 2016–2021,” Tech. Rep.
9 Huawei Technology, “Communication network 2030,” Tech. Rep., 2021.
10 W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and challenges,” IEEE Internet Things J.,
vol. 3, no. 5, pp. 637–646, Oct. 2016.
11 Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial
intelligence with edge computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1738–1762, 2019.
12 J. Park, S. Samarakoon, M. Bennis, and M. Debbah, “Wireless network intelligence at the edge,” Proceedings of the
IEEE, vol. 107, no. 11, pp. 2204–2239, 2019.
13 M. Chen, B. Liang, and M. Dong, “Joint offloading decision and resource allocation for multi-user multi-task mobile
cloud,” in 2016 IEEE International Conference on Communications (ICC), 2016, pp. 1–6.
14 D. Hoang, D. Niyato, and P. Wang, “Optimal admission control policy for mobile cloud computing hotspot with
cloudlet,” in 2012 IEEE Wireless Communications and Networking Conference (WCNC), 2012, pp. 3145–3149.
15 Y. Mao, J. Zhang, S Song, and K. Letaief, “Power-delay tradeoff in multi-user mobile-edge computing systems,” in
2016 IEEE Global Communications Conference (GLOBECOM), 2016, pp. 1–6.
16 F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and computing optimization in wireless powered mobile-edge
computing systems,” IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 1784–1797, Dec. 2018.
17 X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, “Joint computation and communication cooperation for energy-efficient
mobile edge computing,” IEEE Internet Things J., vol. 6, no. 3, pp. 4188–4200, Jun. 2019.
18 G. Zhu, J. Xu, K. Huang, and S. Cui, “Over-the-air computing for wireless data aggregation in massive IoT,” IEEE
Wireless Commu., vol. 28, no. 4, pp. 57–65, Aug. 2021.
 Zhu G, et al. Sci China Inf Sci 19

19 X. Cao, G. Zhu, J. Xu, Z. Wang, and S. Cui, “Optimized power control design for over-the-air federated edge
learning,” IEEE J. Select. Areas Commun., vol. 40, no. 1, pp. 342–358, Jun. 2022.
20 W. Liu, X. Zang, Y. Li, and B. Vucetic, “Over-the-air computation systems: Optimization, analysis and scaling
laws,” IEEE Trans. Wireless Commun., vol. 19, no. 8, pp. 5488–5502, Jan. 2020.
21 X. Cao, G. Zhu, J. Xu, and K. Huang, “Optimized power control for over-the-air computation in fading channels,”
IEEE Trans. Wireless Commu., vol. 19, no. 11, pp. 7498–7513, Aug. 2020.
22 G. Zhu and K. Huang, “MIMO over-the-air computation for high-mobility multimodal sensing,” IEEE Internet
Things J., vol. 6, no. 4, pp. 6089–6103, 2019.
23 W. Fang, Y. Jiang, Y. Shi, Y. Zhou, W. Chen, and K. Letaief, “Over-the-air computation via reconfigurable intelligent
surface,” IEEE Trans. Commun., vol. 69, no. 12, pp. 8612–8626, 2021.
24 W. Zhang, J. Xu, W. Xu, X. You, and W. Fu, “Worst-case design for RIS-aided over-the-air computation with
imperfect CSI,” IEEE Commun. Lett., vol. 26, no. 9, pp. 2136–2140, Jun. 2022.
25 M. Fu, Y. Zhou, Y. Shi, W. Chen, and R. Zhang, “UAV aided over-the-air computation,” IEEE Trans. Wireless
Commun., vol. 21, no. 7, pp. 4909–4924, 2022.
26 F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and S. Buzzi, “Integrated sensing and communications:
Towards dual-functional wireless networks for 6G and beyond,” IEEE J. Select. Areas Commu., vol. 40, no. 6, pp.
1728–1767, Mar. 2022.
27 F. Liu, C. Masouros, A. P. Petropulu, H. Griffiths, and L. Hanzo, “Joint radar and communication design: Ap-
plications, state-of-the-art, and the road ahead,” IEEE Trans. Commu., vol. 68, no. 6, pp. 3834–3862, Feb.
2020.
28 F. Liu, L. Zhou, C. Masouros, A. Li, W. Luo, and A. Petropulu, “Toward dual-functional radar-communication
systems: Optimal waveform design,” IEEE Trans. Signal Process., vol. 66, no. 16, pp. 4264–4279, Aug. 2018.
29 X. Liu, T. Huang, N. Shlezinger, Y. Liu, J. Zhou, and Y. Eldar, “Joint transmit beamforming for multiuser MIMO
communications and MIMO radar,” IEEE Trans Sig. Process., vol. 68, pp. 3929–3944, Jun. 2020.
30 H. Hua, J. Xu, and T. Xiao Han, “Optimal transmit beamforming for integrated sensing and communication,” 2021,
arXiv: 2104.11871. [Online]. Available: https://arxiv.org/pdf/2104.11871.pdf.
31 F. Liu, Y. Liu, A. Li, C. Masouros, and Y. Eldar, “Cramer-Rao bound optimization for joint radar-communication
beamforming,” IEEE Trans. Sig. Process., vol. 70, pp. 240–253, Dec. 2022.
32 Z. Lyu, G. Zhu, and J. Xu, “Joint maneuver and beamforming design for uav-enabled integrated sensing and
communication,” IEEE Trans. Wireless Commun., 2022.
33 Xianxin Song, Jie Xu, Fan Liu, Tony Xiao Han, and Yonina C Eldar, “Intelligent reflecting surface enabled sensing:
Cram\’er-rao bound optimization,” available online: https://arxiv.org/abs/2207.05611, 2022.
34 Xianxin Song, Ding Zhao, Haocheng Hua, Tony Xiao Han, Xun Yang, and Jie Xu, “Joint transmit and reflective
beamforming for IRS-assisted integrated sensing and communication,” in in Proc. 2022 IEEE Wireless Communi-
cations and Networking Conference (WCNC). IEEE, 2022, pp. 189–194.
35 X. Wang, Z. Fei, J. Huang, and H. Yu, “Joint waveform and discrete phase shift design for ris-assisted integrated
sensing and communication system under cramer-rao bound constraint,” IEEE Trans. Veh. Technol., vol. 71, no.
1, pp. 1004–1009, Oct. 2022.
36 W. Shi, W. Xu, X. You, C. Zhao, and K. Wei, “Intelligent reflection enabling technologies for integrated and green
internet-of-everything beyond 5g: Communication, sensing, and security,” IEEE Wireless Commun., pp. 1–8, 2022.
37 X. Li, F. Liu, Z. Zhou, G. Zhu, S. Wang, K. Huang, and Y. Gong, “Integrated sensing and over-the-air computation:
Dual-functional mimo beamforming design,” in 2022 1st International Conference on 6G Networking (6GNet), 2022,
pp. 1–8.
38 Q. Huang, H. Chen, and Q. Zhang, “Joint design of sensing and communication systems for smart homes,” IEEE
Network, vol. 34, no. 6, pp. 191–197, 2020.
39 F. Liu, W. Yuan, C. Masouros, and J. Yuan, “Radar-assisted predictive beamforming for vehicular links: Communi-
cation served by sensing,” IEEE Trans. Wireless Commu., vol. 19, no. 11, pp. 7704–7719, Aug. 2020.
40 W. Xu, Z. Yang, D. Ng, M. Levorato, Y. Eldar, and M. Debbah, “Edge learning for B5G networks with distributed sig-
nal processing: Semantic communication, edge computing, and wireless sensing,” 2022, arXiv:2206.00422. [Online].
Available: https://arxiv.org/pdf/2206.00422.pdf.
41 D. Liu, G. Zhu, J. Zhang, and K. Huang, “Data-importance aware user scheduling for communication-efficient edge
machine learning,” IEEE Trans. Cogn. Commun. Netw., vol. 7, no. 1, pp. 265–278, 2021.
42 S. Wang, Y. Wu, M. Xia, R. Wang, and H. Poor, “Machine intelligence at the edge with learning centric power
allocation,” IEEE Trans. Wireless Commun., vol. 19, no. 11, pp. 7293–7308, 2020.
43 L. Zhou, Y. Hong, S. Wang, R. Han, D. Li, R. Wang, and Q. Hao, “Learning centric wireless resource allocation for
edge computing: Algorithm and experiment,” IEEE Trans. Veh. Technol, vol. 70, no. 1, pp. 1035–1040, 2021.
44 H. Zhang, M. Cisse, Y. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” 2017,
arXiv:1710.09412. [Online]. Available: https://arxiv.org/pdf/1710.09412.pdf.
45 Y. Koda, J. Park, M. Bennis, P. Vepakomma, and R. Raskar, “Airmixml: Over-the-air data mixup for inherently
privacy-preserving edge machine learning,” in 2021 IEEE Global Communications Conference (GLOBECOM), 2021,
pp. 01–06.
46 T. Zhang, S. Wang, G. Li, F. Liu, G. Zhu, and R. Wang, “Accelerating edge intelligence via integrated sensing and
communication,” in ICC 2022 - IEEE International Conference on Communications, 2022, pp. 1586–1592.
47 C. Ding, J. Wang, H. Zhang, M. Lin, and G. Li, “Joint MIMO precoding and computation resource allocation for
 Zhu G, et al. Sci China Inf Sci 20

dual-function radar and communication systems with mobile edge computing,” IEEE J. Select. Areas Commun.,
vol. 40, no. 7, pp. 2085–2102, 2022.
48 Z. Liang, H. Chen, Y. Liu, and F. Chen, “Data sensing and offloading in edge computing networks: TDMA or
NOMA?,” IEEE Trans. Wireless Commun., vol. 21, no. 6, pp. 4497–4508, 2022.
49 Y. Qi, Y. Zhou, Y. Liu, L. Liu, and Z. Pan, “Traffic-aware task offloading based on convergence of communication
and sensing in vehicular edge computing,” IEEE Internet Things J., vol. 8, no. 24, pp. 17762–17777, 2021.
50 F. Roth, N. Bidoul, T. Rosca, M. Dörpinghaus, D. Flandre, A. Ionescu, and G. Fettweis, “Spike-based sensing and
communication for highly energy-efficient sensor edge nodes,” in 2022 2nd IEEE International Symposium on Joint
Communications & Sensing (JC&S), 2022, pp. 1–6.
51 B. Luo, W. Xiao, S. Wang, J. Huang, and L. Tassiulas, “Tackling system and statistical heterogeneity for feder-
ated learning with adaptive client sampling,” in Proceedings of IEEE Conference on Computer Communications
(INFOCOM). IEEE, 2022, pp. 1–10, IEEE.
52 H. Chen, S. Huang, D. Zhang, M. Xiao, M.l Skoglund, and H. V. Poor, “Federated learning over wireless iot networks
with optimized communication and resources,” IEEE Internet Things J., pp. 1–1, 2022.
53 M. Chen, Z. Yang, W. Saad, C. Yin, H V. Poor, and S. Cui, “A joint learning and communications framework for
federated learning over wireless networks,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 269–283, 2020.
54 J. Xu and H. Wang, “Client selection and bandwidth allocation in wireless federated learning networks: A long-term
perspective,” IEEE Trans. on Wireless Commun., vol. 20, no. 2, pp. 1188–1200, 2021.
55 V. Nguyen, S. K. Sharma, T. X. Vu, S. Chatzinotas, and B. Ottersten, “Efficient federated learning algorithm for
resource allocation in wireless IoT networks,” IEEE Internet Things J., vol. 8, no. 5, pp. 3394–3409, 2020.
56 C. T. Dinh, N. H. Tran, M. N. H. Nguyen, C. S. Hong, W. Bao, A. Y. Zomaya, and V. Gramoli, “Federated learning
over wireless networks: Convergence analysis and resource allocation,” IEEE/ACM Trans. Networking, vol. 29, no.
1, pp. 398–409, 2021.
57 Z. Ma, Y. Xu, H. Xu, Z. Meng, L. Huang, and Y. Xue, “Adaptive batch size for federated learning in resource-
constrained edge computing,” IEEE Trans. Mobile Comp., pp. 1–1, 2021.
58 Claudio Ba., Paolo D. L., M. Mattia, and Sergio B., “Lyapunov-based optimization of edge resources for energy-
efficient adaptive federated learning,” TechRxiv preprint, p. 16732651.v1, 2021.
59 B. Luo, W. Xiao, S. Wang, J. Huang, and L. Tassiulas, “Tackling system and statistical heterogeneity for federated
learning with adaptive client sampling,” in Proc. IEEE INFOCOM 2022, 2022, pp. 1739–1748.
60 P. Liu, J. Jiang, G. Zhu, L. Cheng, W. Jiang, W. Luo, Y. Du, and Z. Wang, “Training time minimization for federated
edge learning with optimized gradient quantization and bandwidth allocation,” Frontiers of Information Technology
and Electronic Engineering, vol. 23, no. 8, pp. 1247–1263, 2022.
61 X. Cao, Z. Lyu, G. Zhu, J. Xu, L. Xu, and S. Cui, “An overview on over-the-air federated edge learning,” 2022,
[Online]. Available: http://arXiv.org/pdf/2208.05643.pdf.
62 G. Zhu, Y. Wang, and K. Huang, “Broadband analog aggregation for low-latency federated edge learning,” IEEE
Trans. Wireless Commun., vol. 19, no. 1, pp. 491–506, 2019.
63 N. Zhang and M. Tao, “Gradient statistics aware power control for over-the-air federated learning,” IEEE Trans.
Wireless Commun., vol. 20, no. 8, pp. 5115–5128, Mar. 2021.
64 X. Cao, G. Zhu, J. Xu, and S. Cui, “Transmission power control for over-the-air federated averaging at network
edge,” IEEE J. Select. Areas Commun., pp. 1–1, 2022.
65 K. Yang, T. Jiang, Y. Shi, and Z. Ding, “Federated learning via over-the-air computation,” IEEE Trans. Wireless
Commun., vol. 19, no. 3, pp. 2022–2035, 2020.
66 Y. Sun, S. Zhou, Z. Niu, and D. Gündüz, “Dynamic scheduling for over-the-air federated edge learning with energy
constraints,” IEEE J. Select. Areas Commun., vol. 40, no. 1, pp. 227–242, Nov. 2022.
67 A. Haque, A. Milstein, and L. Fei-Fei, “Illuminating the dark spaces of healthcare with ambient intelligence,” Nature,
vol. 585, no. 7824, pp. 193–202, 2020.
68 D. Liu and O. Simeone, “Privacy for free: Wireless federated learning via uncoded transmission with adaptive power
control,” IEEE J. Select. Areas Commun., vol. 39, no. 1, pp. 170–185, 2020.
69 Hang Liu, Xiaojun Yuan, and Ying-Jun Angela Zhang, “Reconfigurable intelligent surface enabled federated learning:
A unified communication-learning design approach,” IEEE Transactions on Wireless Communications, vol. 20, no.
11, pp. 7595–7609, 2021.
70 Yandong Shi, Yong Zhou, and Yuanming Shi, “Over-the-air decentralized federated learning,” in Proc. 2021 IEEE
International Symposium on Information Theory (ISIT). IEEE, 2021, pp. 455–460.
71 Emre Ozfatura, Stefano Rini, and Deniz Gündüz, “Decentralized SGD with over-the-air computation,” in Proc.
GLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 2020, pp. 1–6.
72 G. Li, S. Wang, J. Li, R. Wang, F. Liu, M. Zhang, X. Peng, and T. X. Han, “Rethinking the tradeoff in integrated
sensing and communication: Recognition accuracy versus communication rate,” 2021, arXiv: 2107.09621. [Online].
Available: https://arxiv.org/pdf/2107.09621.pdf.
73 P. Liu, G. Zhu, S. Wang, W. Jiang, W. Luo, H. V. Poor, and S. Cui, “Toward Ambient Intelligence: Federated Edge
Learning with Task-Oriented Sensing, Computation, and Communication Integration,” 2022, arXiv:2206.05949.
[Online]. Available: https://arxiv.org/pdf/2206.05949.pdf.
74 T. Zhang, S. Wang, G. Li, F. Liu, G. Zhu, and R. Wang, “Accelerating edge intelligence via integrated sensing and
communication,” 2021, arXiv: 2107.09574. [Online]. Available: https://arxiv.org/pdf/2107.09574.pdf.
75 Y. Cui, F. Liu, X. Jing, and J. Mu, “Integrating sensing and communications for ubiquitous IoT: Applications,
 Zhu G, et al. Sci China Inf Sci 21

trends, and challenges,” IEEE Network, vol. 35, no. 5, pp. 158–167, 2021.
76 A. Liu, Z. Huang, M. Li, Y. Wan, W. Li, T. X. Han, C. Liu, R. Du, D. K. P. Tan, J. Lu, et al., “A survey on
fundamental limits of integrated sensing and communication,” IEEE Commun. Surveys Tuts, vol. 24, no. 2, pp.
994–1034, Feb. 2022.
77 X. Li, F. Liu, Z. Zhou, G. Zhu, S. Wang, K. Huang, and Y. Gong, “Integrated sensing, communication,
and computation over-the-air: MIMO beamforming design,” 2022, arXiv: 2201.12581. [Online]. Available:
https://arxiv.org/pdf/2201.12581.pdf.
78 P. Liu, G. Zhu, W. Jiang, W. Luo, J. Xu, and S. Cui, “Vertical federated edge learning with distributed integrated
sensing and communication,” IEEE Commu. Lett., pp. 1–1, Jun. 2022.
79 Jiale Guo, Qiang Liu, and Enqing Chen, “A deep reinforcement learning method for multimodal data fusion in action
recognition,” IEEE Signal Processing Lett., vol. 29, pp. 120–124, Nov. 2022.
80 M. Chen, D. Gündüz, K. Huang, W. Saad, M. Bennis, A. V. Feljan, and H. V. Poor, “Distributed learning in wireless
networks: Recent progress and future challenges,” IEEE J. Select. Areas Commun., vol. 39, no. 12, pp. 3579–3605,
December 2021.
81 W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y. C. Eldar, and M. Debbah, “Edge learning for B5G networks
with distributed signal processing: Semantic communication, edge computing, and wireless sensing,” 2022, arXiv:
2206.00422. [Online]. Available: https://arxiv.org/pdf/2206.00422.pdf.
82 K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conf. on Computer
Vision and Pattern Recognition (CVPR), 2016, pp. 770–778.
83 S. F. Yilmaz, B. Hasircioglu, and D. Gündüz, “Over-the-air ensemble inference with model privacy,” Proc. IEEE
International Symposium on Information Theory (ISIT), pp. 1265–1270, May 2022.
84 K. Yang, Y. Shi, W. Yu, and Z. Ding, “Energy-efficient processing and robust wireless cooperative transmission for
edge inference,” IEEE Internet Things J., vol. 7, pp. 9456–9470, Oct. 2020.
85 S. Hua, K. Zhou, Y.and Yang, Y. Shi, and K. Wang, “Reconfigurable intelligent surface for green edge inference,”
IEEE Trans. Green Commun. Netw., vol. 5, pp. 964–979, Jun. 2021.
86 M. Jankowski, D. Gündüz, and K. Mikolajczyk, “Deep joint source-channel coding for wireless image retrieval,” Proc.
IEEE International Conference on Acoustics, Speech, and Signal Processing. (ICSAAP), pp. 5070–5074, May 2020.
87 M. Jankowski, D. Gündüz, and K. Mikolajczyk, “Wireless image retrieval at the edge,” IEEE J. Select. Areas
Commun., vol. 39, no. 1, pp. 89–100, Jan. 2020.
88 J. Shao, Y. Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck
approach,” IEEE J. Select. Areas Commun., vol. 40, no. 1, pp. 197–211, Jul. 2021.
89 F. Pezone, S. Barbarossa, and P. D. Lorenzo, “Goal-oriented communication for edge learning based on the information
bottleneck,” Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICSAAP), pp.
8832–8836, May 2020.
90 J. Shao, Y. Mao, and J. Zhang, “Task-oriented communication for multi-device cooperative edge inference,” 2021,
arXiv: 2109.00172. [Online]. Available: https://arxiv.org/pdf/2109.00172.pdf.
91 H. Xie, Z. Qin, and G. Y. Li, “Task-oriented multi-user semantic communications for VQA,” IEEE Wireless Commun.
Lett., vol. 11, no. 2, pp. 553–557, Mar. 2022.
92 X. Huang and S. Zhou, “Dynamic compression ratio selection for edge inference systems with hard deadlines,” IEEE
Internet Things J., vol. 7, no. 9, pp. 8800–8810, Sep. 2020.
93 X. Tang, X. Chen, L. Zeng, S. Yu, and L. Chen, “Joint multiuser DNN partitioning and computational resource
allocation for collaborative edge intelligence,” IEEE Internet Things J., vol. 8, no. 12, pp. 9511–9522, Jun. 2021.
94 J. Shao and J. Zhang, “Communication-computation trade-off in resource-constrained edge inference,” IEEE Com-
mun. Mag., vol. 58, no. 12, pp. 20–26, Dec. 2020.
95 Z. Liu, Q. Lan, and K. Huang, “Resource allocation for multiuser edge inference with batching and early exiting,”
2020, arXiv: 2204.05223. [Online]. Available: https://arxiv.org/pdf/2204.05223.pdf.
96 Q. Lan, Q. Zeng, P. Popovski, D. Gündüz, and K. Huang, “Progressive feature transmission for split inference at the
wireless edge,” 2021, arXiv: 2112.07244. [Online]. Available: https://arxiv.org/pdf/2112.07244.pdf.
97 D Wen, X. Jiao, P Liu, G. Zhu, Y. Shi, and K. Huang, “Task-oriented over-the-air computation for multi-device edge
AI,” 2022, arXiv: 2211.01255. [Online]. Available: https://arxiv.org/pdf/2211.01255.pdf.
98 M. Lee, G. Yu, and H. Dai, “Privacy-preserving decentralized inference with graph neural networks in wireless
networks,” 2022, arXiv: 2208.06963. [Online]. Available: https://arxiv.org/pdf/2208.06963.pdf.
99 D. Wen, P. Liu, G. Zhu, Y. Shi, J. Xu, Y. C. Eldar, and S. Cui, “Task-oriented sensing, computation,
and communication integration for multi-device edge AI,” 2022, arXiv: 2207.00969. [Online]. Available:
https://arxiv.org/pdf/2207.00969.pdf.
100 N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” 2000, arXiv: 0004057. [Online].
Available: https://arxiv.org/pdf/physics/0004057.pdf.
101 S. Hua, Y. Zhou, K. Yang, Y. Shi, and K. Wang, “Reconfigurable intelligent surface for green edge inference,” IEEE
Trans. Green Commun. Netw., vol. 5, no. 2, pp. 964–979, 2021.