Source-linked AI summary
Pushing AI to Wireless Network Edge: An Overview on Integrated Sensing, Communication, and Computation towards 6G
Guangxu Zhu, Zhonghao Lyu, Xiang Jiao, Peixi Liu, Mingzhe Chen, Jie Xu, Shuguang Cui, Ping Zhang
TL;DR
Mission-critical AI applications need low-latency, reliable processing of data generated at the network edge, which cloud-centered designs cannot adequately support. This survey reviews task-oriented integrated sensing, communication, and computation for edge intelligence across centralized learning, federated learning, and edge inference. It identifies joint resource management, AirComp, ISAC, and related research opportunities, while noting limitations including straggler bottlenecks and reliance on centralized edge servers in prior work.
Problem
Cloud-based communication-assisted AI must transfer large amounts of edge-generated data to centralized servers, conflicting with mission-critical applications' latency and reliability requirements.
Method
The article surveys task-oriented ISCC for edge intelligence, organizing representative work across centralized edge learning, federated edge learning, and edge inference.
Results
The survey covers joint communication-computation resource management, AirComp, ISAC, and research opportunities across the three edge-intelligence scenarios.
Takeaways & Limitations
Task-oriented ISCC evaluates edge-AI systems through inference or training accuracy, latency, and energy efficiency alongside the successful completion of downstream AI tasks.
Takeaways & Limitations
Prior Air-FEEL studies mainly assume a centralized edge server, while weak-channel devices can bottleneck model aggregation and decentralized settings address server unavailability or unreliability.
Abstract
from arXiv · showhide
Pushing artificial intelligence (AI) from central cloud to network edge has reached board consensus 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 network 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 applications. 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.
1 Introduction
Edge AI moves intelligence toward distributed data sources to meet demanding AI-service requirements, motivating task-oriented integration of sensing, communication, and computation. The survey organizes this integration across centralized learning, federated learning, and edge inference.
- Edge AI: Cloud-based communication-assisted AI struggles with mission-critical applications because large volumes of edge data must be transported for centralized processing.The paper identifies industrial control, VR, AR, and auto-driving as examples.
- Edge AI: 85 Zettabytes of usable data were projected at the network edge in 2021, versus 21 Zettabytes of global data-center traffic.The contrast illustrates the shift of data generation toward the edge.
- Edge AI: Around one millisecond of closed-loop sensing-communication-computation latency is required by applications such as smart industrial Internet, VR, and AR.These services also demand ultra-high reliability and real-time processing.
- Edge AI: Edge computing relocates selected processing and storage from central clouds to distributed nodes near data providers, while edge AI combines these resources with data scattered across the edge.This convergence aims to provide low-latency AI services to mobile terminals.
- ISCC motivation: Traditional networks separate sensing, communication, and computation, although edge AI requires their coupled operation for data acquisition, delivery, and downstream processing.The separation design makes stringent latency, reliability, and capacity demands difficult to accommodate.
- Task-oriented ISCC: Task-oriented ISCC optimizes downstream AI metrics such as inference or training accuracy, latency, and energy efficiency rather than communication throughput alone.The survey applies this perspective to centralized edge learning, federated edge learning, and edge inference.
2 Centralized Edge Learning
Centralized edge learning brings computation close to distributed data but still relies on centralized transmission and processing, creating latency, communication, privacy, and security challenges. Existing work integrates communication, computation, sensing, and learning objectives, while leaving task-oriented ISCC and adaptive data selection as open directions.
- Centralized Edge Learning: Centralized edge learning deploys computation and storage at network-edge servers to address the resource demands of increasingly complex deep-learning models.It brings computation close to data through scaling-up or scaling-out strategies at the edge.
- Centralized Edge Learning: Clients upload collected data to a central edge server, which stores and processes it before returning the learned model.This architecture is simple to deploy and manage for geographically dispersed data, but centralized processing introduces long delays and high transmission demands.
- Challenges: Centralized edge learning can incur catastrophic delay, backbone strain, and central-server computation overhead in latency-sensitive applications.Autonomous driving and VR games are cited as examples of such applications.
- Integrated Communication and Computation: AirMixML combines mixup and AirComp so users transmit modulated private-data signals whose noisy superposition trains models while preserving privacy.Transmit power controls privacy disclosure, while the Dirichlet dispersion ratio controls each worker’s contribution; the framework achieves the same accuracy as raw samples but remains immature for practical use.
- Integrated Sensing and Communication: ISAC unifies sensing and communication through shared wireless signals to support dataset generation and transmission, but introduces additional interference.Prior work jointly optimizes beamforming, time allocation, data offloading, transmit precoding, and computation resources for learning-related objectives.
- Research Opportunities: Open directions include secure uploading against poisoned data, data-importance-aware transmission, and task-oriented joint sensing, communication, and computation resource management.Data importance may change during transmission, and task-oriented ISCC for edge inference remains unexplored.
3 Federated Edge Learning
Federated edge learning (FEEL) must jointly address heterogeneous communication, computation, and data conditions. Air-FEEL and ISAC-based designs improve efficiency by integrating wireless aggregation with sensing and communication, while several task-oriented challenges remain.
- 3.1 Joint Communication and Computation Resource Management in FEEL: Device heterogeneity in radio resources, channel status, and computational capabilities directly affects FEEL accuracy and latency.Synchronized aggregation can make the slowest device determine each round’s communication time.
- 3.2 Over-the-Air Federated Edge Learning (Air-FEEL): Air-FEEL uses over-the-air model or gradient aggregation and can reduce communication latency by a factor approximately equal to the number of devices without significant learning-accuracy loss.Research also examines resource management, device scheduling, and privacy preservation.
- 3.2 Over-the-Air Federated Edge Learning (Air-FEEL): Adaptive transmit-power control reduces aggregation error and can improve learning accuracy or convergence rate in Air-FEEL.Other approaches schedule devices according to signal strength, energy constraints, and computation capabilities.
- 3.2 Over-the-Air Federated Edge Learning (Air-FEEL): Air-FEEL remains vulnerable to stragglers because the weakest-channel device can bottleneck model aggregation; decentralized variants address settings without a reliable edge server.RIS-assisted optimization jointly considers device selection, over-the-air transceiver design, and RIS configuration.
- 3.3 Federated Edge Learning with ISAC: FEEL with ISAC integrates sensing and communication across three levels, from orthogonal resources to shared signal waveforms, with increasing interference-management demands.The third level can improve spectrum, hardware, and information-processing efficiency, while multimodal, dynamic, and task-oriented settings remain open problems.
- 3.4 Research Opportunities: Future FEEL-ISCC systems must handle multimodal sensors, changing sensing environments, and learning tasks whose sensing, communication, and computation modules are coupled, collaborative, or competing.These directions extend beyond studies focused mainly on communication and computation resource optimization.
4 Edge Inference
Edge inference deploys trained models near data sources but faces coupled communication and computation bottlenecks. AirComp and ISCC offer integration mechanisms, while inference characterization, multi-device coordination, and multimodal processing remain open challenges.
- 4 Edge Inference: Edge inference runs trained ML models at the network edge for real-time tasks such as classification, recommendation, and regression.It benefits computation-, storage-, and power-limited devices and delay-sensitive AI services.
- 4 Edge Inference: On-device, on-server, and split inference distribute computation differently, but complex models, limited device resources, wireless transmission, and processing delays complicate deployment.Joint management of sensing, communication, and computation resources is consequently more difficult.
- 4.2 Joint Communication and Computation Resource Management in Edge Inference: Joint optimization studies balance inference accuracy, transmission cost, model split points, computational resources, and network power consumption.Examples include dynamic compression under hard deadlines, split-point selection, and RIS-aided beamforming and task assignment.
- 4.3 AirComp-based Edge Inference: AirComp integrates communication and computation for low-latency edge inference, but over-the-air edge inference remains in its early stage.An initial multi-device system aggregates noisy feature observations to average feature noise and improve inference accuracy.
- 4.4 Edge Inference with ISCC: ISAC-enabled edge inference must characterize inference performance and jointly design sensing, communication, and computation resources.Open directions include fundamental JSCC limits, multi-device or multi-server coordination, task-oriented resource management, and multimodal sensing-data processing.
5 Concluding Remarks
The paper argues that 6G edge intelligence requires moving AI toward the network edge and replacing separated sensing, communication, and computation with task-oriented ISCC. It surveys representative techniques across centralized learning, FEEL, and edge inference, then identifies future research directions.
- 5 Concluding Remarks: 6G is expected to extend network functionality from connecting people and things to connecting intelligence and delivering pervasive AI services at the edge.This shift supports the AIoT vision.
- 5 Concluding Remarks: Separated sensing, communication, and computation cannot meet stringent latency, reliability, and capacity demands, motivating task-oriented ISCC.The survey covers centralized edge learning, federated edge learning, and edge inference, and highlights future research directions.