PaperlayerView exact revision on arXiv

Source-linked AI summary

Edge Intelligence: The Confluence of Edge Computing and Artificial Intelligence

Shuiguang Deng, Hailiang Zhao, Weijia Fang, Jianwei Yin, Schahram Dustdar, Albert Y. Zomaya

arXiv:1909.00560v2cs.NIcs.DC

TL;DR

Edge-generated data and AI advances create a need to integrate Edge Computing with AI. The paper classifies Edge Intelligence into AI for edge and AI on edge and organizes the field through a research road-map. It presents AI for edge as AI-assisted optimization of Edge Computing and AI on edge as model training and inference at the edge, with QoE spanning performance, cost, privacy, efficiency, and reliability.

  • Problem

    The paper addresses how to integrate Edge Computing and AI across a broad, diverse, and partly unclear research field involving massive distributed edge data and constrained optimization.

  • Method

    The paper establishes a broader perspective by classifying Edge Intelligence into AI for edge and AI on edge and arranging their research concerns in a hierarchical road-map.

  • Results

    The paper identifies AI for edge with AI-assisted solutions to Edge Computing problems and AI on edge with training and inference of AI models on edge.

  • Takeaways & Limitations

    The classification and road-map provide background for understanding concepts and potential future research initiatives in Edge Intelligence.

Abstract

from arXiv · show

Along with the rapid developments in communication technologies and the surge in the use of mobile devices, a brand-new computation paradigm, Edge Computing, is surging in popularity. Meanwhile, Artificial Intelligence (AI) applications are thriving with the breakthroughs in deep learning and the many improvements in hardware architectures. Billions of data bytes, generated at the network edge, put massive demands on data processing and structural optimization. Thus, there exists a strong demand to integrate Edge Computing and AI, which gives birth to Edge Intelligence. In this paper, we divide Edge Intelligence into AI for edge (Intelligence-enabled Edge Computing) and AI on edge (Artificial Intelligence on Edge). The former focuses on providing more optimal solutions to key problems in Edge Computing with the help of popular and effective AI technologies while the latter studies how to carry out the entire process of building AI models, i.e., model training and inference, on the edge. This paper provides insights into this new inter-disciplinary field from a broader perspective. It discusses the core concepts and the research road-map, which should provide the necessary background for potential future research initiatives in Edge Intelligence.

I. INTRODUCTION

Edge Intelligence emerges from the convergence of expanding edge-generated data, Edge Computing, and rapidly advancing AI. The paper addresses this broad and diverse field by distinguishing AI for edge from AI on edge and presenting a broader research perspective.

  • Motivation: Edge Computing pushes computation and communication resources from the cloud toward network edges to reduce unnecessary communication latency and congestion.The motivation includes geographically distributed mobile and IoT data and the difficulty of offloading massive volumes to the cloud.
  • Motivation: AI is increasingly used to process large volumes of data and extract insights for better decisions.Deep Neural Networks and related deep-learning architectures are identified as prominent AI technologies.
  • Motivation: The integration of Edge Computing and AI creates Edge Intelligence, a complex field spanning interwoven concepts and technologies.The paper emphasizes that Edge Intelligence is more than a simple combination of its constituent areas.
  • Classification: The paper distinguishes Edge Intelligence into AI for edge and AI on edge.AI for edge targets constrained Edge Computing optimization, whereas AI on edge runs AI-model training and inference through device-edge-cloud synergy.
  • Paper scope: The paper provides a broader perspective, simple classification, research road-map, state-of-the-art discussion, and grand challenges for both directions.This structure responds to the diverse literature and unresolved concepts identified by the authors.

II. THE RELATIONS BETWEEN EDGE COMPUTING AND AI

AI and Edge Computing reinforce one another: AI supplies methods for Edge Computing’s optimization needs, while Edge Computing supplies distributed data, scenarios, and platforms for AI applications.

  • Interactive relationship: The paper characterizes the relationship as interactive: AI can expand Edge Computing’s potential, while Edge Computing can expand AI’s applicability.The two fields are presented as mutually reinforcing rather than merely adjacent.
  • AI supports Edge Computing: AI provides technologies and methods for Edge Computing’s resource-allocation and system-efficiency problems.The paper identifies allocation challenges involving CPU cycles, access jurisdiction, radio frequency, and bandwidth.
  • Edge Computing supports AI: Edge Computing provides AI with geographically distributed data and application scenarios such as vehicles, autonomous driving, smart homes, and smart cities.These scenarios facilitate the movement of AI applications from theory toward practice.

III. RESEARCH ROAD-MAP OF EDGE INTELLIGENCE

The research road-map separates AI for edge and AI on edge while organizing their shared objective around application-dependent Quality of Experience. QoE jointly considers performance, cost, privacy, efficiency, and reliability.

  • Road-map structure: The road-map logically separates AI for edge on the left from AI on edge on the right.AI for edge is organized bottom-up through Topology, Content, and Service, while AI on edge is decomposed top-down through Model Adaptation, Framework Design, and Processor Acceleration.
  • Quality of Experience: Quality of Experience is the shared optimization objective placed at the top of the road-map.The objective applies collectively to both Edge Intelligence directions.
  • Quality of Experience: QoE is application-dependent and jointly considers performance, cost, privacy or security, efficiency, and reliability.These criteria capture both outcome quality and resource, protection, and operational concerns.
  • QoE criteria: Cost includes computation, communication, and energy consumption, with latency also treated as a target for reduction.Energy is especially important because mobile devices have limited battery capacity.
  • QoE criteria: Privacy-preserving distributed training motivates Federated Learning, while efficiency methods include model compression, conditional computation, and algorithm asynchronization.These methods address data leakage concerns and the overhead of training and inference.
  • QoE criteria: Reliability is particularly important for AI on edge because distributed, synchronized training and inference can be disrupted by wireless congestion and incomplete model transfers.The paper ties reliability to avoiding failures during prescribed operating periods.

B. A Recapitulation of IEC

AI for edge, termed Intelligence-enabled Edge Computing, applies AI to complex Edge Computing problems. Its research concerns are organized into Topology, Content, and Service, covering deployment, provisioning, placement, caching, offloading, migration, and mobility.

  • B. A Recapitulation of IEC: AI for edge is called Intelligence-enabled Edge Computing and uses AI to solve complex learning, planning, and decision-making problems.Its bottom-up organization categorizes Edge Computing concerns into Topology, Content, and Service.
  • Topology: Topology covers Orchestration of Edge Sites and Wireless Networking.An edge site is defined as a micro data center with applications deployed and attached to a Small-cell Base Station.
  • Topology: Learning-based mobility prediction can improve transmit power and QoE satisfaction relative to benchmarks.The cited example uses a conceptor-based echo state network to learn user mobility patterns.
  • Content: Content includes Data Provisioning, Service Provisioning, Service Placement, Service Composition, and Service Caching.These concerns address how resources and complex services are provided and deployed across cloud and edge infrastructure.
  • Service: Service focuses on Computation Offloading, User Profile Migration, and Mobility Management.Computation offloading includes edge-server selection and frequency-spectrum allocation for balancing computational and communication resources.

C. A Recapitulation of AIE

AI on edge (AIE) covers training and inference at the network edge through model adaptation, framework design, and processor acceleration. The section emphasizes distributed training, collaborative inference, and methods that reduce communication or computation demands.

  • AI on edge studies training and inference of AI models on the network edge, organized into Model Adaptation, Framework Design, and Processor Acceleration.
  • Framework Design: Framework design develops edge training and inference architectures without modifying existing AI models.
  • Framework Design: Distributed training can split data or models across cloud, edge servers, and mobile devices, including student networks trained on private end-device data.
  • Framework Design: Federated Learning trains local models on multiple clients and averages their gradients into a global model without centralizing private data.
  • Framework Design: Model inference commonly uses model splitting, compression, input filtering, or early exit to distribute computation while respecting inference-accuracy constraints.
  • Model Adaptation: Model adaptation improves existing frameworks through compression, conditional computation, asynchronous aggregation, and decentralization to reduce communication or computation.

3) Processor Acceleration:

Processor acceleration optimizes DNN computation through specialized hardware structures and execution strategies. In the broader AI-for-edge workflow, AI technologies are applied after defining and mathematically formulating edge optimization problems.

  • Processor Acceleration: Processor acceleration optimizes DNN structure and multiply-and-accumulate operations through specialized instructions, parallel computation, and near-data processing.
  • AI-for-edge workflow: The paper presents AI utilization for performance optimization in the Mobile Edge Computing environment.
  • AI-for-edge workflow: AI-for-edge optimization begins by identifying objectives, decision variables, and constraints before constructing a mathematical model.
  • AI-for-edge workflow: Performance optimization may combine task execution delay, transmission delay, and task-dropping cost, with binary or partial task decisions.
  • AI-for-edge scope: The paper focuses its AI-for-edge discussion on Wireless Networking, Service Placement, Service Caching, and Computation Offloading.

1) Wireless Networking:

AI for wireless networking applies learning across data acquisition, radio-resource management, and signal encoding, while related work also addresses power, placement, and caching decisions. These approaches target edge-network optimization using communication-aware or learning-based designs.

  • Wireless Networking: Learning-driven Communication applies machine learning across multiple access, radio resource management, and signal encoding.
  • Wireless Networking: Learning-driven multiple access exploits wireless-channel characteristics for functional computation, including over-the-air aggregation of Federated Learning updates.
  • Wireless Networking: Learning-driven radio resource management allocates resources according to transmitted-data value, exemplified by importance-aware retransmission.
  • Wireless Networking: Learning-driven signal encoding jointly optimizes feature extraction, source coding, and channel encoding through schemes such as Hybrid Federated Distillation.
  • Wireless Networking: Other AI-for-wireless studies use Graph Neural Networks for scalable power control in K-user interference channels.
  • Related edge optimization: Related edge-optimization work models service placement as a Markov Decision Process and uses reinforcement learning or multi-armed bandits for placement decisions.
  • Related edge optimization: Service caching addresses fetch-cache decisions for popular content under dynamic prices and user requests.

3) Computation Offloading:

Computation offloading transfers resource-intensive tasks from constrained mobile devices to edge or cloud systems while allocating computing and communication resources. AI methods are widely applied to this optimization, but model restrictions and constrained search spaces remain major challenges.

  • Computation Offloading: Computation offloading transfers resource-intensive tasks from resource-limited mobile devices to edge or cloud infrastructure.
  • Computation Offloading: Offloading requires allocating resources such as CPU cycles and channel bandwidth, making optimization-capable AI methods especially relevant.
  • Computation Offloading: Q-learning and DQN are prominent approaches for offloading optimization, including energy-harvesting IoT devices in multi-server MEC systems.
  • Computation Offloading: DQN-based offloading can maximize utility combining data-sharing gains, task-dropping penalties, energy consumption, and computation delay.
  • Computation Offloading: Other approaches predict task popularity with LSTM networks and jointly optimize offloading, computation-resource allocation, and caching decisions.
  • Challenges: AI-based edge optimization requires restricted mathematical models and finite or discretized state and action spaces, which can degrade performance.

2) Algorithm Deployment:

AI for edge faces difficult deployment conditions: many problems are combinatorial and computationally expensive, while edge resources constrain efficiency. AI on edge therefore organizes deployment around adaptation, framework design, and processor acceleration.

  • Combinatorial and NP-hard formulations create high computational complexity for many state-of-the-art edge problems.
  • AI-for-edge solutions commonly rely on iterative learning, but online deployment introduces additional challenges.
  • Resource-constrained edges require balancing solution optimality against efficiency under dynamically changing QoE requirements and network resources.
  • AI-on-edge research is divided into Model Adaptation, Framework Design, and Processor Acceleration.Training frameworks include Federated Learning and Knowledge Distillation, while inference frameworks include Model Splitting and Model Partitioning.
  • Model Adaptation is presented through methods, approaches, and technologies summarized in Fig. 3.

A. State of the Art

Model compression is a central AI-on-edge direction, using compression and partitioning techniques to fit models and communication to constrained edge environments. However, existing work is often insufficiently connected to specific edge applications and platforms.

  • 1) Model Compression:: Model Compression reduces memory and channel occupation by exploiting sparsity in gradients and weights.Approaches include quantization, dimensionality reduction, pruning, component sharing, precision downgrading, SVD, Huffman coding, and PCA.
  • 1) Model Compression:: Federated Learning prioritizes communication efficiency because offline devices and network congestion can prevent global-model updates.Structured updates learn in a restricted lower-dimensional space, while sketched updates compress the model before uploading.
  • 1) Model Compression:: Secure aggregation can tolerate up to 33.3% of participating devices failing to complete the protocol.The cited protocol is designed for high-dimensional data and is described as robust.
  • 1) Model Compression:: Partitioned DNNs support collaborative inference by executing a quantized first part on edge devices and a full-precision second part in the cloud.
  • 1) Model Compression:: ProtoNN compresses k-nearest-neighbor learning by representing the training set with a small number of prototypes.The prototypes are learned through Stochastic Neighborhood Compression and lower-dimensional projection.
  • 1) Model Compression:: Model Compression is active and easy to implement, but state-of-the-art work is usually not tied to specific edge-computing application scenarios.The paper identifies opportunities to construct edge platforms and hardware.

2) Conditional Computation:

Conditional computation selectively avoids less important calculations, while asynchronous aggregation and decentralization address communication failures and privacy risks in federated edge learning.

  • 2) Conditional Computation:: Conditional computation selectively turns off unimportant calculations through component sharing, shutoff, input filtering, early exit, and result caching.It can be viewed as block-wise dropout, with ranking, selection, or confidence thresholds controlling computation.
  • 2) Conditional Computation:: The same selective-computation idea can support Federated Learning participant selection by excluding less valuable participants from global-model aggregation.
  • 2) Conditional Computation:: Algorithm Asynchronization aggregates local models asynchronously to address failures in model uploading and downloading caused by wireless congestion.Peer-to-peer exchange of weights and gradients, including randomized gossip, is another approach to reducing wireless-channel concurrency.
  • 2) Conditional Computation:: Thorough Decentralization removes the central aggregator to reduce privacy leakage from model updates.Blockchain and game-theoretical approaches are identified as mechanisms for total decentralization.
  • 2) Conditional Computation:: The paper frames data availability, model selection, and coordination mechanism as the grand challenges for AI on edge.

1) Data Availability:

AI-on-edge deployment depends on obtaining usable training data, selecting models and infrastructure under resource limits, and coordinating heterogeneous devices across cloud, edge, and device tiers.

  • 1) Data Availability:: Data availability is a fundamental challenge because raw training data may require incentives from mobile users before it can support training and inference.Data from heterogeneous end devices can also complicate robust communication-protocol design.
  • 1) Data Availability:: Model selection must jointly address learning-accuracy thresholds, model scale, training frameworks, and accelerator architectures under limited resources.The paper links model selection with resource allocation and management.
  • 1) Data Availability:: Heterogeneous computing and communication resources can cause the same Model Adaptation method to produce different learning results across device clusters.The paper therefore emphasizes compatibility and coordination among heterogeneous edge devices.
  • 1) Data Availability:: The paper classifies Edge Intelligence into AI for edge and AI on edge and presents both research road-maps hierarchically.AI for edge addresses Edge Computing concerns with AI, while AI on edge covers model training and inference at the edge.
Loading 1909.00560v2…