PaperlayerView exact revision on arXiv

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

arXiv:2004.14850v1cs.DCcs.AIcs.NI

TL;DR

Edge intelligence is considered a missing element in 5G and a prospective enabler for 6G, while its design principles and technological advances need an overarching account. This white paper discusses edge-AI challenges and potential solutions, identifies research questions, and presents a roadmap toward intelligent edge AI.

  • Problem

    Edge intelligence is considered a missing element in 5G and a prospective 6G enabler, while its design principles and technological advances need an overarching account.

    Source: §1 · p. 2

  • Method

    The paper discusses edge-AI challenges and potential solutions in emerging 6G, and frames key research questions and a roadmap.

    Source: §1 · p. 4

  • Results

    The white paper identifies edge-AI challenges and research questions and presents a roadmap toward intelligent edge AI.

    Source: §1 · p. 4

  • Takeaways & Limitations

    The vision is for edge nodes to learn and share knowledge, collectively providing added-value or optimized services.

    Source: §1 · p. 3

Abstract

from arXiv · show

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 presents Edge Intelligence (EI) as a key enabler for 6G, arguing that growing AI demands and data generation at the network edge call for intelligence beyond centralized cloud computing. It surveys the concepts, challenges, enabling technologies, and research roadmap for developing intelligent edge services.

Introduction

  • EI is identified as a key missing element in 5G and a likely key enabler of future 6G performance, functions, and services.The paper envisions AI features at individual edge nodes and clusters that learn progressively and may share knowledge to provide collective or optimized services.Source: §1 · pp. 2–3
  • Hundreds of ZB of data are generated at the network edge, motivating an extension of centralized cloud computing toward edge resources.Edge computing moves some service-specific processing and storage to nodes near data providers and users; edge and fog computing are used interchangeably for execution outside the central cloud.Source: §1 · pp. 2–3
  • EI analyzes data at or near where it is generated and used, reducing latency, costs, and security risks through edge-deployed services and functions.The paper describes an ETSI MEC host and platform supporting edge applications through standardized APIs, including access to base-station information and network-slicing support.Source: §1 · p. 3
  • Current AI solutions are resource- and energy-hungry and time-consuming, while increasingly stringent near-real-time demands expose hardware limits and motivate photonic and memory-integrated computing approaches.The paper discusses nanophotonic computation for DNN matrix multiplication, in-memory computing to reduce processor-memory data movement, and ACAP architectures integrating CPUs, AI/DSP engines, and programmable logic.Source: §1 · pp. 3–4
  • Divergent software tools and device constraints leave responsibilities predefined at design time, hindering liquid software and relocation of intelligence across the network.The paper argues that without software able to flow between devices, intelligence placement must be decided during design because computations cannot be easily relocated without preparation.Source: §1 · p. 4
  • The white paper examines edge AI challenges and potential solutions, then presents a vision, key enablers, research questions, and a roadmap toward intelligent edge AI.Its stated coverage includes related work, the edge-AI vision, 6G challenges and enablers, research questions, and a roadmap.Source: §1 · p. 4

2. RELATED WORK

Related work surveys edge intelligence architectures, technologies, and challenges, while other papers examine use cases, platforms, theory, and the roles of AI and edge computing. Together, these works frame the technical and organizational questions surrounding edge intelligence.

2. RELATED WORK

  • Zhou et al. and Xu et al. [28] survey edge AI, covering DNNs, edge training and inference architectures, emerging technologies, challenges, and future directions.Their reviews focus on AI running at the network edge, especially DNNs for supervised learning.Source: §2 · p. 5
  • Rausch and Dustdar categorize edge-intelligence use cases as public, private, or intersecting and raise questions about ownership, providers, telecommunications, government, and engineering.Examples include smart public spaces, personal health assistants, predictive maintenance, and autonomous vehicles.Source: §2 · p. 5
  • OpenEI addresses edge data-analysis challenges with lightweight intelligent processing and data sharing, alongside low-power AI tools and broader edge-computing models and initiatives.The broader examples include a layered service model, a summary of 75 initiatives [7], and ECCE’s industrial reference architecture, technology stacks, and pathfinders.Source: §2 · p. 5
  • Park et al. analyze neural-network splits and theoretical enablers for distributed, low-latency, reliable machine learning at the wireless edge, with case studies toward 5G and beyond.Their work links edge learning to mission-critical applications and intelligent devices.Source: §2 · p. 5
  • Lovén et al. distinguish Edge for AI from AI for Edge and outline a secure, privacy-aware platform; Deng et al. distinguish AI for Edge from AI on Edge.Deng et al. describe AI on Edge as running the AI-model lifecycle through device-edge-cloud synergy, subject to performance, cost, privacy, reliability, and efficiency constraints.Source: §2 · pp. 5–6

3. VISION FOR THE 2030S EDGE-DRIVEN ARTIFICIAL INTELLIGENCE

The paper envisions 6G edge-driven AI as enabling an “Intelligent Internet of Intelligent Things,” with distributed intelligence supporting future services and network requirements.

3. VISION FOR THE 2030S EDGE-DRIVEN ARTIFICIAL INTELLIGENCE

  • 6G edge AI can support the shift to an “Intelligent Internet of Intelligent Things,” while base stations integrate computing and communications and may create new opportunities for telecom operators.Edge intelligence is presented as a means to improve network efficiency and enable added-value services; future applications also demand trust, security, reliability, and resilience.Source: §3 · pp. 7–8
  • Edge computing faces heterogeneous data sources and constraints in hardware, environment, availability, reliability, and energy; performance, cost, security, efficiency, and reliability are key indicators.The paper cites Mohan on edge-integration challenges and proposed solutions.Source: §3 · p. 8
  • Zhou et al. define six edge-intelligence levels; the paper extends this to seven generic-AI levels, from cloud-based operation to fully on-device training and inference.The levels distinguish where training and inference occur, including cloud-edge, in-edge, and edge-device cooperation.Source: §3 · pp. 8–9
  • Edge nodes can act as local clouds and clusters can share contextual knowledge and computational and communication loads; orchestration should flexibly support coexisting AI-for-Edge and AI-on-Edge services.The paper proposes an ontology for 6G connectivity to shape possible combinations of edge microservices.Source: §3 · pp. 9–10

DATA

  • Edge Intelligence uses advanced communications and AI for data collection, aggregation, fusion, processing, distribution, and services at the edge, with learning, inference, and control in static and dynamic environments.The passage describes learning, inference, and control as added-value capabilities.Source: DATA · p. 11

COOPERATIVE

  • AI algorithms on geographically distant heterogeneous platforms can jointly solve a problem, with distance imposing latency requirements.Source: COOPERATIVE · p. 11

REAL-TIME

  • Localized AI/ML functions operate with constrained computation resources and usually strict real-time requirements.Source: REAL-TIME · p. 11

REQUIREMENTS

  • Intelligent computing should be available wherever and whenever users need it, meeting requirements for computing power, latency, energy, cost, mobility, and service reliability.Source: REQUIREMENTS · p. 11

4. CHALLENGES AND KEY ENABLERS · EDGE INFRASTRUCTURE SOLUTIONS · EDGE PLATFORM ORCHESTRATION

Edge intelligence’s training and inference bring technical challenges that require new enablers. The paper highlights distributed edge-platform control and workload-aware infrastructure planning as central concerns.

4. CHALLENGES AND KEY ENABLERS

  • Realizing edge intelligence for both training and inference raises technical challenges that motivate identifying new theoretical and technical enablers.The paper contrasts these challenges with traditional centralized AI systems.Source: §4 · p. 12

EDGE INFRASTRUCTURE SOLUTIONS

  • The MEC reference architecture targets an open, multi-vendor edge platform, with system- and host-level management under a centralized orchestrator’s authority over resources.Platform and host components have partial autonomy within their domains and provide feedback to the orchestrator.Source: EDGE INFRASTRUCTURE SOLUTIONS · p. 12
  • This control model challenges real-time reactivity, low-latency multi-tenant service, data routing and aggregation, and system-information delivery in dynamic IoT environments.Using system-wide resource, performance, delivery, QoS, and QoE data for AI-based learning and optimization remains largely unexplored.Source: EDGE INFRASTRUCTURE SOLUTIONS · p. 12
  • Edge infrastructure spans two-tier MEC [12], hierarchical fog, cloudlets, mist, and mobile computing, making autonomous local management especially important and challenging.These approaches place computation at different levels, from cloud-linked edge platforms to data-producing devices and user equipment.Source: EDGE INFRASTRUCTURE SOLUTIONS · pp. 12–13

EDGE PLATFORM ORCHESTRATION

  • AI-based orchestration suits opportunistic, large-scale IoT edge systems, where resource efficiency and QoE require many data sources and complex analysis.Centralized algorithms can be difficult to design, deploy, maintain, and evaluate, while distributed edge execution aligns with the underlying architecture.Source: EDGE PLATFORM ORCHESTRATION · p. 13
  • Infrastructure allocation must determine deployment locations and capabilities while respecting budgets and device limits, using historical data and online or predicted workloads to plan.Deployments may follow existing network infrastructure, such as wireless access points, or use low-resource Fog units.Source: EDGE PLATFORM ORCHESTRATION · p. 13

DATA AND NETWORK MANAGEMENT · INTERSECTION OF EDGE AND DEVICE INTELLIGENCE

Edge intelligence depends on reliable, context-appropriate data and adaptive network management, while distributed edge-device applications must coordinate resources and AI across changing environments. The sections identify data-processing techniques, network modeling, and autonomous agents as responses to these challenges.

DATA AND NETWORK MANAGEMENT

  • Small-data generalization, duplicate and anomalous records, heterogeneous sensory data, and cross-network inconsistencies can all undermine edge-intelligence performance.The data may combine audio, video, haptic, spatial, temporal, and stochastic characteristics, making fusion a further challenge.Source: DATA AND NETWORK MANAGEMENT · p. 13
  • Changing network states and requirements demand adaptable models; sharing data fractions and trained models instead of raw data can reduce communication payloads.The paper calls for defining data and model provenance and lifecycle, and comparing models’ fit to current contexts.Source: DATA AND NETWORK MANAGEMENT · p. 13
  • Clustering, anomaly detection, and down-sampling can refine large edge datasets and help manage data inconsistencies.For small data, the paper also proposes increased sampling frequency, synthetic data generation with generalized adversarial networks, incremental learning, knowledge distillation, and transfer learning.Source: DATA AND NETWORK MANAGEMENT · p. 14
  • Feature extraction and split learning across modalities can address heterogeneous data, while generalized adversarial networks may restore consistency with synthetic data.These approaches complement the paper’s proposed small-data methods for developing higher-quality models over time.Source: DATA AND NETWORK MANAGEMENT · p. 14
  • Rare network failures, imbalanced data, heavy-tailed traffic, and AI-based flow or path selection complicate network optimization and prediction.The paper proposes network simulation using a digital twin to produce balanced data, adding localized weather and seasonal foliage information, and understanding network behavior rather than treating it as a black box.Source: DATA AND NETWORK MANAGEMENT · p. 14

INTERSECTION OF EDGE AND DEVICE INTELLIGENCE

  • Mobile-edge applications partition software across devices, edge, and cloud, sharing refined data and resources to conserve device energy and respond to user mobility.Uniform interfaces, large-scale orchestration, and control latency are challenges, motivating lightweight device-level AI for autonomy and self-* capabilities.Source: INTERSECTION OF EDGE AND DEVICE… · p. 14
  • Distributed AI agents and multi-agent systems support autonomy, reactivity, adaptivity, learning, code mobility, and collaboration, including on resource-constrained IoT devices.The paper identifies these capabilities as benefits of the classical distributed AI paradigm.Source: INTERSECTION OF EDGE AND DEVICE… · p. 14
  • In cloud-edge-device continuums, agents represent entities, coordinate resource sharing, and control or monitor systems; evolving from reactive to deliberative, learning, proactive agents remains open.Potential complementary techniques include swarm intelligence, game theory, and genetic algorithms.Source: INTERSECTION OF EDGE AND DEVICE… · p. 15

SOFTWARE DEVELOPMENT FOR EDGE · REAL-TIME REQUIREMENTS AND ONLINE LEARNING

6G edge intelligence depends on automated, policy-aware software deployment and management, while real-time AI requires redesigning the full feedback cycle rather than relying on lower latency and higher bandwidth alone.

SOFTWARE DEVELOPMENT FOR EDGE

  • Edge software management must automate image building, deployment, lifecycle handling, and application-specific service, data, network access, and QoE under system policies.Packages may involve multiple stakeholders; online negotiation and orchestration must also address SLAs, access rights, billing, security, and resource and information sharing.Source: SOFTWARE DEVELOPMENT FOR EDGE · p. 15
  • Microservices offer on-demand modular deployment as an alternative to resource-intensive, multi-gigabyte VM images, while unikernels may offer better isolation at similar image sizes to containers.Containers encapsulate individual microservices; unikernels can run as virtual machines or on bare metal.Source: SOFTWARE DEVELOPMENT FOR EDGE · p. 15
  • Edge DevOps must adapt automated CI/CD and AI workflow frameworks such as Kubeflow9 and MLFlow10 to edge data sources and federated cloud environments.Managing versioning and deployment at scale becomes challenging as application-specific units and workflows expand system and package management and performance monitoring.Source: SOFTWARE DEVELOPMENT FOR EDGE · p. 15

REAL-TIME REQUIREMENTS AND ONLINE LEARNING

  • Real-time feedback matters for AI applications including robotics, self-driving cars, traffic and logistics management, and telepresence, virtual, and augmented reality.These applications are included in 6G verticals and application areas.Source: REAL-TIME REQUIREMENTS AND ONLINE… · pp. 15–16
  • Real-time edge AI must balance pre-trained and online-learned models, distribute and reuse models efficiently, and make decisions using knowledge across models and data sources.The feedback cycle includes data collection, model training, and returning actions to applications, so network latency and bandwidth improvements alone are insufficient.Source: REAL-TIME REQUIREMENTS AND ONLINE… · p. 16
  • Transfer learning and knowledge distillation can use preacquired intelligence to reduce retraining latency, while distillation and pruning can reduce model size for faster inference.Adapting data dissemination and model training to network changes and reducing inference complexity address short response-time needs; reinforcement learning and communication-control-ML codesign address data and network dynamics [20].Source: REAL-TIME REQUIREMENTS AND ONLINE… · p. 16

DEVELOPING DISTRIBUTEDLY TRAINED ALGORITHMS

Most mission-critical and privacy-sensitive edge applications require online distributed training, while device constraints, communication bottlenecks, and changing networks shape the algorithms needed for reliable, scalable edge intelligence.

DEVELOPING DISTRIBUTEDLY TRAINED ALGORITHMS

  • Most mission-critical and privacy-sensitive edge applications require online distributed training, because training affects end-to-end latency, inference reliability, and system scalability.Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 16
  • Although larger models and frequent coordination can improve inference accuracy and reliability, device limits and uncertain links demand energy-efficient algorithms with resource scheduling and data or model compression.With limited data, devices may exchange raw data, model parameters, or inferred outputs and decisions to improve reliability and robustness.Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 16
  • Training can parallelize data across mobile edge servers or split large models across devices, while federated learning, knowledge distillation, and transfer learning suit differing data, privacy, and model needs [22] [23] [25].Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 16
  • Communication efficiency can exploit wireless uplink-downlink capacity asymmetry through joint distillation and federated learning, and use pruning or coded and quantized transmission to reduce communication and storage demands [26].Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 17
  • Because static algorithms may perform poorly as networks change, distributed training can use cold-start retraining or continual knowledge acquisition through transfer, online, and reinforcement learning.Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 17
  • Meta-distribution, risk-management, and extreme-value-theory frameworks can analyze generalization error, while differential privacy, rate-distortion, and mean-field control tools support low-latency, secure, scalable algorithms.Source: DEVELOPING DISTRIBUTEDLY TRAINED… · p. 17

SECURITY AND PRIVACY · END-USER ASPECTS

Edge intelligence must secure data and processing across edge-cloud links while protecting privacy and preserving user QoE through adaptation to context, mobility, and changing resource conditions. The sections highlight distributed security and privacy-preserving training alongside user-centered resource management, migration, incentives, and trust.

SECURITY AND PRIVACY

  • Edge-cloud transmission exposes data and processing outcomes to attacks and tampering, while edge locality and AI can support security and privacy measures such as face anonymization.The paper also gives replacing people with straw figures as an example when an application needs only counts or detection that someone has fallen.Source: SECURITY AND PRIVACY · p. 17
  • Lightweight distributed security must support authentication, access control, model and data integrity, platform verification, and secure routing across trusted and malicious edge nodes.Source: SECURITY AND PRIVACY · p. 17
  • Federated learning keeps original datasets on their generating devices or nodes while sharing model parameters, with differential privacy, homomorphic encryption, and secure multiparty computation as further protections.Directly sharing privacy-sensitive datasets across edge nodes risks leakage and may conflict with requirements such as GDPR.Source: SECURITY AND PRIVACY · p. 17

END-USER ASPECTS

  • Maintaining user QoE requires understanding behavioral context and reacting to local connectivity, latency, bandwidth, transmission, execution, and integration needs through online, on-demand edge-resource adaptation.Adaptation can propagate horizontally and vertically across the deployment.Source: END-USER ASPECTS · pp. 17–18
  • User mobility requires migrating virtualized components, data, and stateful applications across edge deployments while minimizing handover latency, and multi-tenant resource sharing raises privacy and security concerns.Source: END-USER ASPECTS · p. 18
  • Mist computing needs participation incentives such as micropayments, data-ownership and sharing policies, and GDPR-compliant privacy protection; digital twins or cognitive software agents can represent users, with trust needed for cooperation.Source: END-USER ASPECTS · p. 18

PRICING AND SHARING MECHANISM

The section proposes sharing mobile-edge communication, caching, computation, and learning resources through an intelligent framework, with dynamic pricing and distributed mechanisms to support allocation, incentives, and resource utilization.

PRICING AND SHARING MECHANISM

  • A virtual pool for sharing communication, caching, computation, and learning resources is proposed to improve utilization and meet 6G application QoE and QoS requirements.Applications named include Tactile Internet, virtual reality, and autonomous driving.Source: PRICING AND SHARING MECHANISM · p. 18
  • Dynamic pricing would model edge devices as agents that broker or purchase resources, while market equilibrium and distributed learning guide prices and allocations for differing QoE and QoS requirements.The model also treats 3C resources and knowledge as tradable.Source: PRICING AND SHARING MECHANISM · p. 18
  • Smart contracts and distributed ledgers are proposed to support fair, secure resource sharing and revenue dissemination, alongside incentives and lightweight consensus protocols.The passage also identifies recording prices and distributing revenue according to proof of work among distributed edge devices as important.Source: PRICING AND SHARING MECHANISM · p. 18

End-user aspects

  • End-user aspects mention understanding user context from large-scale behavior patterns and cognitive digital twins and software agents.Source: End-user aspects · p. 19

5. CORE RESEARCH QUESTIONS

Demand for responsiveness, privacy, and context-awareness is pushing intelligence to the edge; the paper frames the core research questions for developing 6G Edge Intelligence across AI for Edge and AI on Edge services.

EDGE INFRASTRUCTURE

  • Edge infrastructure research must define flexible application architectures, edge-component taxonomies, collaboration across devices and network topologies, and protocols for outer-network and in-network intelligence.The questions include NSF-level and inter-NSF intelligence.Source: EDGE INFRASTRUCTURE · p. 20

SYSTEM PLATFORM AND STACK

  • System-platform research must establish software development, quality assurance, and testing practices, alongside DevOps, cooperative decentralized diagnostics, and formal verification.Source: SYSTEM PLATFORM AND STACK · p. 20

DATA AND NETWORK MANAGEMENT; INTERSECTION OF EDGE AND DEVICE INTELLIGENCE

  • Data and network management must support online learning with non-stationary data, federated methods, energy efficiency, agent-based cognition, and distributed computing.At the edge-device intelligence intersection, questions include lightweight autonomous AI, cognitive agents, and collaborative optimization using swarm intelligence, game theory, and genetic algorithms.Source: DATA AND NETWORK MANAGEMENT · p. 20, INTERSECTION OF EDGE AND DEVICE… · p. 20

SOFTWARE DEVELOPMENT OF EDGE; REAL-TIME REQUIREMENTS AND ONLINE LEARNING

  • Edge software research must enable interoperable, reconfigurable agents and continuous deployment, while supporting virtualization and software that moves between nodes.Real-time and online-learning questions include transfer learning, knowledge distillation, reinforcement learning, communication-control-ML codesign, rapid model adaptation to network changes, and lower inference complexity.Source: SOFTWARE DEVELOPMENT OF EDGE · p. 20, REAL-TIME REQUIREMENTS AND ONLINE… · p. 20

DISTRIBUTED TRAINING, ALGORITHMIC DESIGN AND DEPLOYMENT

  • Research must address accelerators, economically sustainable edge resources, energy-computation tradeoffs, and communication-efficient, reliable, scalable distributed training, inference, and control across heterogeneous platforms.The questions also include factorizing ML algorithms to run partly across heterogeneous computing resources.Source: DISTRIBUTED TRAINING, ALGORITHMIC… · p. 21

SECURITY, PRIVACY, AND PORTABILITY

  • Security, privacy, and portability questions include protecting data, models, and protocols, fail-safe operation during network or node failures, scalable data provenance, and satisfying user and regulatory privacy concerns.Source: SECURITY, PRIVACY, AND PORTABILITY · p. 21

END-USER CONCERNS

  • End-user research must address QoE standardization, resource sharing, mobility, stateful application migration, social and regulatory concerns, and relationships among edge stakeholders and network actors.Source: END-USER CONCERNS · p. 21

PRICING AND SHARING MECHANISM

  • Pricing and sharing questions include virtualizing 3C-L resources, market-equilibrium sharing, dynamic pricing, resource trading and revenue distribution, and incentives.Source: PRICING AND SHARING MECHANISM · p. 21

6. PROSPECTIVE USE CASES

The section presents edge intelligence as an enabler for applications in autonomous mobility, smart spaces, environmental sensing, mobile XR, and collaborative manufacturing.

EDGE INTELLIGENCE FOR AUTONOMOUS DRIVING

  • As platoon traffic grows, congestion may prevent DSRC or C-V2X from meeting 99.99% reliability; edge intelligence enables dynamic spectrum access for fast, reliable vehicular data processing.Database-oriented systems can specify available bands, with edge processing and decision units co-located at base stations or roadside units.Source: EDGE INTELLIGENCE FOR AUTONOMOUS DRIVING · pp. 22–23

EDGE INTELLIGENCE FOR SMART SPACES

  • Smart spaces need edge-native solutions to handle latency-sensitive services, privacy-sensitive data, and large data volumes, including hospital data processing on premises and real-time surveillance recognition.Regulations such as GDPR may require hospital data to be stored and processed within the premises.Source: EDGE INTELLIGENCE FOR SMART SPACES · p. 23

EDGE INTELLIGENCE FOR ENVIRONMENTAL SENSING

  • Environmental sensing needs dense sensor data collection and real-time processing; edge-native processing can reduce network load and parallelize computation.AI can help choose sensor locations and trajectories, calibrate low-cost sensors, and determine where computation should occur.Source: EDGE INTELLIGENCE FOR ENVIRONMENTAL… · p. 23

EDGE INTELLIGENCE FOR MOBILE XR

  • Edge intelligence can support resource- and energy-constrained mobile XR through task segmentation, computation offloading, and learning-model sharing to reduce device energy demands and end-to-end latency.Visuohaptic XR generates substantial real-time data at the network edge and requires low latency.Source: EDGE INTELLIGENCE FOR MOBILE XR · p. 23

NEXT GENERATION COBOTS (COLLABORATIVE ROBOTS) IN MANUFACTURING

  • Future cobots will monitor machine health, navigate factory floors, adapt tasks, and collaborate as fleets, requiring real-time data, low-latency communication, and coordination with MES and factory clouds.Edge intelligence can provide fine-grained cobot control while coordinating broader production goals with back-end systems.Source: NEXT GENERATION COBOTS (COLLABORATIVE… · pp. 23–24

7. ROADMAP TO EDGE INTELLIGENCE

The roadmap envisions Edge Intelligence developing over the next decade through advances in edge hardware, software, and increasingly distributed, adaptive AI systems.

Roadmap to the Edge Intelligence

  • The deployment of a new generation of edge intelligence systems, applications, and services is expected during the next ten years, with first deployments of Edge AI indicated for 2020 and 2030.Source: §7 · pp. 24–25
  • Edge hardware must become economically viable at scale, while new AI accelerators target better performance with lower energy use and cost, and nanophotonics may support complex operations or storage.The roadmap also identifies specialized edge devices capable of AI computation and nanophotonic circuits.Source: §7 · pp. 24–25
  • Software advances must support distribution, automation, intelligent orchestration, and security as edge systems and applications develop.The paper also calls for proposals that use the distributed edge architecture to protect data and intellectual property and ensure user privacy.Source: §7 · p. 24
  • Edge AI progresses from pre-trained models toward edge nodes that learn and share models, with distributed training and data processing enabling real-time decisions.The paper expects early deployments to use pre-trained models, later combining them with online learning; the roadmap also depicts AI features at each edge node and distributed algorithms.Source: §7 · pp. 24–25
  • Communications and learning will be combined through learning-driven design principles, alongside approaches to secure information and preserve user privacy.The roadmap depicts learning-driven communication and real-time training as part of the developing system.Source: §7 · pp. 24–25

Related papers

Loading 2004.14850v1…