Source-linked AI summary
6G White Paper on Edge Intelligence
Ella Peltonen, Mehdi Bennis, Michele Capobianco, Merouane Debbah, Aaron Ding, Felipe Gil-Castiñeira, Marko Jurmu, Teemu Karvonen, Markus Kelanti, Adrian Kliks, Teemu Leppänen, Lauri Lovén, Tommi Mikkonen, Ashwin Rao, Sumudu Samarakoon, Kari Seppänen, Paweł Sroka, Sasu Tarkoma, Tingting Yang
TL;DR
The white paper addresses how edge intelligence can support future 6G performance, functions, services, and distributed AI applications. It develops a vision, identifies challenges and enablers across edge infrastructure, data, software, algorithms, and deployment, and proposes research questions and a roadmap toward intelligent edge services.
Problem
Future 6G networks need edge intelligence to support new performance requirements, functions, services, and distributed AI applications across heterogeneous, constrained environments.
Method
The paper provides a vision and roadmap by examining edge infrastructure, data and network management, edge software, distributed training, hardware, security, privacy, pricing, and end-user aspects.
Results
The paper identifies key edge-intelligence challenges, enablers, and research questions, and envisions a transition from the Internet of Things to an Intelligent Internet of Intelligent Things.
Takeaways & Limitations
6G edge intelligence is framed as highly distributed AI in which edge nodes progressively learn, share knowledge, and provide optimized or new services.
Abstract
from arXiv · showhide
In this white paper we provide a vision for 6G Edge Intelligence. Moving towards 5G and beyond the future 6G networks, intelligent solutions utilizing data-driven machine learning and artificial intelligence become crucial for several real-world applications including but not limited to, more efficient manufacturing, novel personal smart device environments and experiences, urban computing and autonomous traffic settings. We present edge computing along with other 6G enablers as a key component to establish the future 2030 intelligent Internet technologies as shown in this series of 6G White Papers. In this white paper, we focus in the domains of edge computing infrastructure and platforms, data and edge network management, software development for edge, and real-time and distributed training of ML/AI algorithms, along with security, privacy, pricing, and end-user aspects. We discuss the key enablers and challenges and identify the key research questions for the development of the Intelligent Edge services. As a main outcome of this white paper, we envision a transition from Internet of Things to Intelligent Internet of Intelligent Things and provide a roadmap for development of 6G Intelligent Edge.
1. INTRODUCTION
The paper frames edge intelligence as a key missing element for 5G and a likely enabler of 6G services, while identifying hardware, software, data, and distributed-intelligence challenges. It proposes a roadmap addressing these challenges across edge infrastructure, AI development, and deployment.
- Edge Intelligence is presented as a key missing element in 5G and a likely enabling factor for future 6G performance, functions, and services.
- The paper connects growing AI use and edge-generated data with the need to process intelligence near users, devices, and monitored systems.Applications discussed include personal assistants, surveillance, smart cities, autonomous vehicles, and Industry 4.0.
- Edge intelligence can distribute AI capabilities across nodes and clusters so they progressively learn and share knowledge for optimized or new services.The paper links this evolution to highly distributed AI in 6G and intelligent autonomous systems requiring safety in human-machine cooperation.
- Current edge AI hardware and software remain immature, while AI solutions are resource-, energy-, and time-intensive.The paper identifies nanophotonics and in-memory computing as approaches to accelerate computation and reduce processor-memory transfer bottlenecks.
- Divergent edge software and design-time placement of intelligence limit the relocation of computation across devices, motivating liquid software for future 6G networks.
- The white paper surveys edge AI challenges and potential solutions, then proposes a roadmap toward intelligent edge AI.Its structure covers related work, the edge-AI vision, challenges and enablers, research questions, and a roadmap.
2. RELATED WORK
Related work surveys edge-intelligence architectures, use cases, platforms, theoretical foundations, and research roadmaps. Across these works, the field spans AI running at the edge, AI improving edge orchestration, and unresolved questions about ownership, architecture, and system integration.
- Existing surveys review the motivation, architectures, frameworks, and technologies for training and inference of deep learning models at the network edge.
- Prior work categorizes edge-intelligence use cases as public, private, and intersecting, while leaving ownership and governance of the future edge fabric unresolved.Open questions include the roles of utilities, telecommunications, governments, and the public in shaping architectures and engineering practices.
- OpenEI and related libraries provide lightweight or low-power support for edge intelligence, data sharing, and AI computation under constrained resources.
- Theoretical work emphasizes distributed, low-latency, reliable machine learning at the wireless edge and analyzes neural-network split tradeoffs.
- Position papers distinguish Edge for AI from AI for Edge and identify communication, control, security, privacy, and applications as key intersection areas.Related architectures support distributed learning, inference, and decision-making by edge-native AI agents.
3. VISION FOR THE 2030S EDGE-DRIVEN ARTIFICIAL INTELLIGENCE
The paper envisions 6G Edge Intelligence as a transition toward an Intelligent Internet of Intelligent Things, combining distributed AI, edge computing, and coordinated communication and computation resources. It organizes this vision around AI for Edge and AI on Edge services, flexible edge architectures, and requirements spanning performance, security, efficiency, reliability, latency, energy, and cost.
- 2030s vision: The vision extends the Internet of Things toward an Intelligent Internet of Intelligent Things supported by 6G communication and edge-driven AI.The paper frames this evolution as necessary for more reliable, efficient, resilient, and secure intelligent services.
- 2030s vision: 6G telecom infrastructure can tightly integrate communication and computing, with base stations serving as natural deployment sites for edge intelligence.This integration is presented as a potential opportunity for telecom and tower operators to increase the value of their offerings.
- Design requirements: Future edge services must balance performance, cost, security, efficiency, reliability, latency, energy consumption, mobility, and service requirements across heterogeneous distributed resources.The paper identifies heterogeneous platforms, constrained real-time AI functions, and user requirements as central architectural considerations.
- Edge-intelligence architecture: The paper defines seven edge-intelligence levels spanning cloud-based, edge-cooperative, in-edge, device-cooperative, and fully on-device training and inference.The levels vary according to where training and inference occur and how data is offloaded or retained locally.
- AI for Edge and AI on Edge: AI for Edge optimizes telecom infrastructure and edge-network life cycles, while AI on Edge provides application-oriented distributed AI services on edge platforms.The paper distinguishes these service categories while emphasizing that both can be distributed across edge nodes and clusters.
- Design requirements: The proposed architecture uses flexible orchestration to reduce dependencies between AI for Edge and AI on Edge and support combinations of edge microservices.An ontology for 6G connectivity is suggested to shape these possible service combinations.
4. CHALLENGES AND KEY ENABLERS
6G edge intelligence must address technical challenges in training and inference that differ from centralized AI. The paper organizes these challenges around distributed edge architectures, data-driven management, and new theoretical and technical enablers.
- Edge intelligence requires new enablers because training and inference at the edge pose challenges unlike traditional centralized AI.
- Edge infrastructure: The ETSI MEC reference architecture uses distributed system- and host-level management, while a centralized orchestrator retains authority over system resources.Platform- and host-level components have partial autonomy and provide operational feedback to orchestrators.
- Architectural approaches: Edge architectures span two-tier MEC, hierarchical fog computing, and cloudlets with deployable physical capacity and application-movement costs.
- Challenges: The opportunistic, physically extensive IoT environment motivates AI for orchestration, resource optimization, and quality-of-experience management.
DATA AND NETWORK MANAGEMENT
Data and network management are central to edge intelligence because edge systems rely on small, changing, and heterogeneous data under dynamic operating conditions. The paper highlights preprocessing, adaptation, and distributed AI techniques as responses.
- Data management: Edge intelligence often relies on small data, making reliable generalization to unseen data a critical challenge.Data availability, accessibility, and type determine the feasibility of edge intelligence; duplicates and anomalies also require identification.
- Network management: Changing network states and tight response times require AI models to adapt, while sharing data or trained models instead of raw data can reduce communication payloads.
- Data management: Clustering, anomaly detection, and down-sampling can refine large edge-device datasets and support efficient, reliable, and generalized models.
- Network management: Rare failures create highly imbalanced learning data, while self-similar traffic produces heavy-tailed distributions that anomaly detection must handle.
- Edge-device continuum: Application location strongly affects real-time reactivity and adaptivity, motivating mobile cloud, mobile edge, and mobile fog platforms despite limited low-end capabilities.
- Edge-device continuum: Software agents and multi-agent systems provide autonomy, reactivity, adaptivity, learning, code mobility, and collaboration across cloud-edge-device systems.Increasing agents from reactive operation toward deliberative, cognitive, learning, and proactive behavior remains an open question.
SOFTWARE DEVELOPMENT FOR EDGE
Edge software development relies on virtualization and automated application lifecycle management, but distributed modularity and real-time feedback increase orchestration complexity. The paper identifies transfer learning, distillation, and pruning as latency-oriented enablers.
- Software platforms: Edge applications are packaged, automatically built, deployed on virtualized hosts, and managed throughout their lifecycle by edge-system components.
- Software platforms: Microservices decompose applications into independently deployable components, reducing the deployment burden of monolithic virtual-machine images.Virtual-machine images are typically several gigabytes, making them resource-consuming to deploy and move across edge platforms.
- Software platforms: CI/CD automates microservice versioning and deployment, but many application-specific units and workflows expand management and performance-monitoring scope.
- Real-time operation: Real-time AI applications such as robotics, autonomous vehicles, logistics, and extended reality require rapid feedback.
- Real-time operation: Reducing latency requires redesigning the full feedback cycle because data collection, model training, and action definition also consume time.
- Real-time operation: Transfer learning, knowledge distillation, and model pruning can reduce retraining latency, model size, and inference cost at the edge.
DEVELOPING DISTRIBUTEDLY TRAINED ALGORITHMS
Distributed edge training must balance inference reliability and scalability against device, communication, energy, storage, and privacy constraints. The paper proposes parallelization, compression, adaptive learning, and lightweight security mechanisms as key enablers.
- Training requirements: Online distributed training is needed for many mission-critical and privacy-concerned applications, because training affects end-to-end latency, inference reliability, and scalability.
- Training requirements: Large models and frequent coordination can improve inference accuracy and reliability but conflict with energy, memory, storage, and privacy constraints.
- Distributed algorithms: Data and model parallelization can address limited device power and memory while accommodating different privacy requirements.
- Distributed algorithms: Pruning, coded transmission, quantization, cold-start mechanisms, continual learning, and reinforcement learning address communication limits and changing networks.Distributed training algorithms must also provide latency, reliability, and scalability guarantees.
- Security: Edge-cloud data transmission exposes information and services to penetration, interception, theft, and denial-of-service attacks.
- Security: Lightweight distributed security should support authentication, access control, integrity, mutual platform verification, secure routing, and trust topologies.
END-USER ASPECTS
Edge intelligence must maintain user quality of experience while adapting to context, mobility, connectivity, latency, bandwidth, and application requirements. This requires online, on-demand resource adaptation and migration across dynamic, multi-tenant edge deployments.
- END-USER ASPECTS: Edge systems target required QoE through improved connectivity, application execution, and adaptation to dynamic environments and user mobility.User context, including location awareness, must be understood from large-scale behavioral patterns.
- END-USER ASPECTS: Online and on-demand adaptation of local edge resources must respond to connectivity, latency, bandwidth, data transmission, and application execution requirements.Adaptation propagates horizontally and vertically across the deployment.
- END-USER ASPECTS: User mobility creates challenges for moving virtualized components and migrating data and stateful applications while minimizing handover latency.Multi-tenant sharing of dynamic edge resources also raises privacy and security concerns.
PRICING AND SHARING MECHANISM
Future 6G edge devices are expected to share communication, caching, computation, and learning resources, but frameworks for coordinating this sharing remain immature. The paper points to dynamic pricing and market-based mechanisms as economic enablers for tradable resources and knowledge.
- PRICING AND SHARING MECHANISM: AI-powered mobile edge devices can share communication, caching, computation, and learning resources to satisfy QoE and QoS requirements for 6G applications.The paper identifies intelligent 3C-L resource sharing as an area that remains in its infancy.
- PRICING AND SHARING MECHANISM: Dynamic pricing can model 3C-L resource sharing by treating mobile edge devices as intelligent agents that price, purchase, and consume resources or services.The proposed perspective supports resource and knowledge exchange among autonomous agents.
- PRICING AND SHARING MECHANISM: A market-equilibrium approach is proposed for making 3C-L resources and knowledge tradable.The supplied passages frame this as an economic sharing model rather than a completed solution.
5. CORE RESEARCH QUESTIONS
The paper formulates core research questions spanning edge architectures, software engineering, distributed learning, lightweight AI, virtualization, security, mobility, resource economics, and energy-aware computation. These questions define a broad roadmap for making edge intelligence flexible, collaborative, adaptive, and deployable.
- 5. CORE RESEARCH QUESTIONS: Research questions address flexible edge architectures, component taxonomies, network impacts, software development, quality assurance, diagnostics, and formal verification.They also ask how decentralized edge entities can cooperate in debugging.
- 5. CORE RESEARCH QUESTIONS: The paper asks how to organize online and federated learning with non-stationary data, improve energy efficiency, enable agent-based cognition, and distribute computing across resources.These questions connect learning organization with distributed resource use.
- 5. CORE RESEARCH QUESTIONS: Further questions concern lightweight AI, cognitive agents, vertical and horizontal cooperation, swarm intelligence, game theory, genetic algorithms, interoperability, reconfigurability, and liquid software.The goal is to support adaptable software agents and continuous deployment across nodes.
- 5. CORE RESEARCH QUESTIONS: The roadmap includes transfer learning, knowledge distillation, reinforcement learning, communication-control-ML codesign, rapid model adaptation, and lower inference complexity.These questions focus on adapting models and reducing the computational burden of edge inference.
- 5. CORE RESEARCH QUESTIONS: Resource questions cover accelerators, economically sustainable ML hardware, energy-computation trade-offs, and distributed training and inference.The paper explicitly includes hardware acceleration, GPUs, computational power, storage capacity, and power consumption.
- 5. CORE RESEARCH QUESTIONS: Security, fail-safe operation, data provenance, privacy, QoE, resource sharing, mobility, stateful migration, regulation, incentives, virtualization, and dynamic pricing remain open questions.The scope extends from technical reliability and privacy to social, regulatory, and economic relationships among edge stakeholders.
6. PROSPECTIVE USE CASES
The paper highlights Edge Intelligence use cases in personal and urban environments, autonomous driving, smart spaces, environmental sensing, extended reality, and collaborative manufacturing. These applications combine latency-sensitive or computation-intensive workloads with distributed data collection, control, and coordination needs.
- 6. PROSPECTIVE USE CASES: Edge Intelligence is positioned for personal computing, urban computing, and manufacturing applications requiring efficient, safe, secure, robust, and resilient wireless networking.The paper uses these domains to motivate the Intelligent Internet of Intelligent Things.
- 6. PROSPECTIVE USE CASES: Autonomous vehicle platooning coordinates groups of driverless cars, creating communication demands among platoon members and with other devices.The platoon cars are assumed to use Cooperative Adaptive Cruise Control for mobility control.
- 6. PROSPECTIVE USE CASES: Dynamic spectrum access can address prospective medium congestion in platoon communications and stringent reliability requirements by offloading some data to other bands.The figure presents Edge Intelligence in the context of vehicular dynamic spectrum access in platoons.
- 6. PROSPECTIVE USE CASES: Smart spaces require latency-sensitive services, privacy-sensitive information exchange, and processing for large data volumes generated by devices such as surveillance cameras.Examples include smart homes, campuses, offices, and hospitals.
- 6. PROSPECTIVE USE CASES: Environmental sensing needs dense sensor deployments and real-time processing to obtain high spatial and temporal resolution air-quality data.Relevant measurements include humidity, temperature, particulate matter, and gaseous pollutants.
- 6. PROSPECTIVE USE CASES: Mobile XR generates massive real-time edge data while supporting computation-intensive visuohaptic interaction with real and virtual elements.The paper identifies VR, AR, and MR as forms of extended reality relevant to 6G services.
- 6. PROSPECTIVE USE CASES: Collaborative robots require real-time data, low-latency communication, and tight coordination with manufacturing execution systems and private clouds.Edge Intelligence can support fine-grained cobot control and coordination of larger production goals.
7. ROADMAP TO EDGE INTELLIGENCE
The roadmap toward Edge Intelligence spans successive deployments, from pre-trained models and specialized edge devices to distributed, real-time learning and learning-driven communications. It also identifies hardware, software, security, privacy, energy, and cost challenges that must be addressed for 6G edge systems.
- Roadmap: The roadmap covers the evolution from first Edge AI deployments toward a new generation of edge intelligence systems, applications, and services through 2030.The transition is presented as a sequence of technological steps over the next ten years.
- Software and networks: Software must advance in distribution, automation, intelligent orchestration, and security as edge intelligence systems grow more complex.These software advances are identified alongside hardware evolution as central implementation challenges.
- Roadmap: Initial deployments use pre-trained AI models at the edge, followed by edge nodes that learn and share models with other nodes.The roadmap depicts pre-trained models, per-node AI features, and model sharing as stages of Edge AI development.
- Distributed intelligence: Future systems will distribute both training and data processing, combine pre-trained with online-learned models, and define actions from collected information in real time.The roadmap links distributed algorithms and real-time training with increasingly adaptive edge intelligence.
- Long-term directions: Learning-driven communications, nanophotonic technologies, and distributed architectures are identified as longer-term directions for efficient computation, secure data, and user privacy.The white paper connects these directions with complex operations, information storage, intellectual-property protection, and privacy preservation.
- Hardware: Specialized edge devices and new AI accelerator ASICs are expected to support increasingly demanding computation while reducing energy consumption and cost.Hardware must become economically viable at scale while supporting DNNs and larger volumes of data.