PaperlayerView exact revision on arXiv

Source-linked AI summary

Edge-Cloud Polarization and Collaboration: A Comprehensive Survey for AI

Jiangchao Yao, Shengyu Zhang, Yang Yao, Feng Wang, Jianxin Ma, Jianwei Zhang, Yunfei Chu, Luo Ji, Kunyang Jia, Tao Shen, Anpeng Wu, Fengda Zhang, Ziqi Tan, Kun Kuang, Chao Wu, Fei Wu, Jingren Zhou, Hongxia Yang

arXiv:2111.06061v3cs.LGcs.AI

TL;DR

Edge-cloud AI lacks a systematic account spanning polarized computing paradigms and their collaboration mechanisms. This survey reviews cloud, edge, and collaborative AI architectures and identifies collaboration types across privacy, efficiency, and personalization goals.

  • Problem

    Edge-cloud AI lacks a systematic review spanning diverse algorithmic areas, heterogeneous edge conditions, and collaboration between cloud and edge models.

  • Method

    The survey systematically reviews cloud AI, edge AI, and edge-cloud collaboration architectures, including federated learning, personalization, compression, and advanced edge topics.

  • Results

    The survey establishes a complete scope of cloud, edge, and collaborative AI and identifies collaboration types ranging from privacy-primary to efficiency-primary personalization.

  • Takeaways & Limitations

    Cloud-edge collaboration can be organized around distinct privacy, efficiency, and personalization objectives across AI computing paradigms.

  • Takeaways & Limitations

    Open-source edge datasets remain scarce because fine-grained edge data are rarely transmitted, and cloud-based datasets cannot fully capture edge characteristics.

Abstract

from arXiv · show

Influenced by the great success of deep learning via cloud computing and the rapid development of edge chips, research in artificial intelligence (AI) has shifted to both of the computing paradigms, i.e., cloud computing and edge computing. In recent years, we have witnessed significant progress in developing more advanced AI models on cloud servers that surpass traditional deep learning models owing to model innovations (e.g., Transformers, Pretrained families), explosion of training data and soaring computing capabilities. However, edge computing, especially edge and cloud collaborative computing, are still in its infancy to announce their success due to the resource-constrained IoT scenarios with very limited algorithms deployed. In this survey, we conduct a systematic review for both cloud and edge AI. Specifically, we are the first to set up the collaborative learning mechanism for cloud and edge modeling with a thorough review of the architectures that enable such mechanism. We also discuss potentials and practical experiences of some on-going advanced edge AI topics including pretraining models, graph neural networks and reinforcement learning. Finally, we discuss the promising directions and challenges in this field.

1 INTRODUCTION

Cloud and edge computing polarize AI research around cloud-scale model capabilities and edge constraints, while real-world systems increasingly require their collaboration. This survey addresses the fragmented edge-cloud landscape by reviewing cloud AI, edge AI, collaboration architectures, advanced edge topics, and future challenges.

  • Cloud and edge polarization: Cloud AI emphasizes model generalization, robustness, fairness, and generation across computer vision, natural language processing, and industrial applications,,,,.Cloud computing provides adequate resources for cost-efficient application computing, whereas edge computing targets low-latency services.
  • Edge AI challenges: Edge AI remains immature because constrained IoT resources make processing speed, throughput, latency, memory, energy, power consumption, and thermal performance critical limitations,.Model-parameter storage and querying can be time-consuming and energy-intensive during inference.
  • Edge AI: Edge AI responds to these constraints by using edge resources and lightweight models, with MobileNet reducing parameters from 29.3 million to 4.2 million and computations eightfold while losing 1% accuracy.EfficientNet is introduced to scale CNNs through network depth, width, and input dimensions.
  • Motivation: Edge-cloud collaboration is necessary because complete real-world systems span cloud and edge from physical infrastructure through algorithmic design.Early examples include federated learning, Taobao EdgeRec, and CloudCNN–EdgeCNN cooperation.
  • Survey scope: The survey organizes a systematic edge-cloud review spanning cloud AI, edge AI, collaboration architectures, advanced deployed-edge topics such as pretraining models, graph neural networks, and reinforcement learning, plus future directions.The need for this scope arises because diversity across algorithmic areas prevents a complete systematic review, while cloud-only large models remain impractical for many edge settings.

2 CLOUD AI

Cloud AI attributes AI’s recent advances, including performance surpassing humans on some open-source competition benchmarks,, to large-scale distributed cloud clusters that accelerate model training. The section reviews computing hardware relevant to image- and text-based applications.

  • 2 CLOUD AI: Cloud computing’s large-scale distributed clusters greatly accelerate AI-model training, underpinning the cloud AI paradigm.The paper identifies this cloud-computing benefit as a main reason for AI’s recent success.
  • 2 CLOUD AI: AI has recently achieved tremendous development and surpassed human performance on some open-source competition benchmarks,.
  • 2 CLOUD AI: The section reviews computing hardware for real-world AI applications involving image and text information carriers.

2.1 AI Computing Hardware in Cloud Computing

Cloud AI’s success depends on hardware, software, and customer-friendly resource management such as virtualization –. This section highlights mainstream computing hardware that has supported AI’s growth, while noting CPU limits for large-scale DNN training.

  • Hardware overview: Mainstream cloud AI hardware, together with software and virtualization technologies –, has played a crucial role in AI’s growth.The survey refers readers to for a complete analysis and focuses here on computing hardware from leading companies.
  • CPU: High-performance CPUs accelerate compute-intensive AI tasks, but general CPU servers are slow for large-scale DNN training.Task-scheduling speedups or combinations with other modules, are therefore commonly used.
  • GPU: GPUs shifted from graphics toward DNN acceleration, enabling AlexNet’s ImageNet success with an NVIDIA GeForce GTX 580.Subsequent NVIDIA products, including Tesla K80 and Tesla P100, further improved computational performance.
  • TPU: Google’s TPU is a dedicated integrated circuit for DNN computation, progressing from TPUv1’s high-bandwidth loop to TPUv2’s cost-effective mixed-precision training.TPUv3 improved performance over TPUv2 for larger-scale network architectures, and Google launched TPUv4 in 2021.

2.2 Computer Vision

Computer vision derives semantics from visual inputs and is reviewed here through image recognition, object detection, and image segmentation. The section also notes progress in super-resolution, image restoration, and image generation.

  • Computer Vision: Computer vision covers image recognition, object detection, and image segmentation as core tasks for interpreting images, videos, and other visual inputs.Image recognition identifies objects, actions, and other elements; detection locates objects; segmentation separates images into recognizable segments.
  • Image Recognition: Image recognition progressed from the neocognitron and CNN-based handwritten-character recognition to AlexNet and residual learning, which scaled convolutional layers to 152 and won multiple 2015 vision tasks.LeCun et al. achieved an error rate below 1% for handwritten digital character recognition, while AlexNet won the ImageNet competition.
  • Object Detection: Object detectors – have improved, but their deep convolution operations increase computing costs and reduce detection speed, motivating research toward real-time detection.The section distinguishes one-stage and two-stage detectors as recent approaches with noticeable improvements.
  • Image Segmentation: Image segmentation requires substantial pixel-labeling effort, while IIC and AC, use mutual-information-based methods to produce or optimize pixel-category probabilities.IIC uses mutual-information-based clustering to output a semantic segmentation probability map; AC uses an autoregressive model and maximizes mutual information across two orderings.
  • Other Computer Vision Tasks: Other advancing computer-vision tasks include super-resolution, image restoration, and image generation [64].DALL-E [64] trains a 12-billion-parameter autoregressive transformer for zero-shot text-to-image generation and achieves significant generalization.

2.3 Natural Language Processing

This section surveys three challenging NLP subareas: machine translation, question answering, and text generation, alongside their representative approaches and applications. It highlights neural and pretrained-model advances, including multimodal generation.

  • Scope: NLP encompasses broad tasks including tagging, named entity recognition, question answering, and machine translation, while this section focuses on three recently challenging subareas, [66].Other part-of-speech tagging and text categorization tasks are deferred to the cited literature, [66].
  • Machine Translation: Machine translation progresses from rule-based and statistical methods to neural approaches using recurrent models, bidirectional LSTM encoder-decoding, attention, and BERT-based architectures –.BERT-based neural architectures achieve state-of-the-art performance across multiple translation benchmarks by increasing model capacity with large-scale corpora.
  • Question Answering: Question answering supports systems such as Microsoft Windows and Apple Siri, typically combining query processing, answer selection, and fine-grained ranking across text, images, and videos –.Research models question-answer pairs and negative sampling to capture implicit matching between multiple objects and answers.
  • Text Generation: Text generation has emerged with large-scale pretrained models such as GPT, spanning poetry and story generation and exploring GAN- and VAE-style mechanisms for more human-like output, –.The discussion also covers knowledge-related generation and cross-domain multimodal tasks such as visual question answering and reading comprehension, [39].

2.4 Web Services

Web services rely on cloud AI models for recommendation, search, and e-advertisement, with research spanning retrieval, personalization, and computational advertising.

  • Web Services: Recommendation, search, and e-advertisement are established web-service paradigms supported by widely explored cloud AI models in enterprises such as Google, Amazon, and Microsoft.
  • Web Search: Web search retrieves relevant objects from vast candidate sets, while image and video search require stronger feature extraction under domain bias and noise.
  • Recommender Systems: Recommender systems are indispensable web-service infrastructure, evolving from collaborative filtering and matrix factorization toward deep learning and sequential modeling.
  • Advertisement: Computational advertising combines advertising, marketing, and computer science to model web data and user preferences while accounting for cost, revenue, and targeting.It also seeks automatic optimization in complex marketing environments while maintaining advertising efficiency relative to traditional advertising.

3 EDGE AI

Mobile AI applications have expanded, but cloud-based inference raises privacy and network-dependence concerns; edge inference addresses these issues by placing models partly or fully on devices for local prediction.

  • Mobile AI supports applications including virtual assistants and personalized recommendation.
  • Cloud-based AI requires uploading data, which may compromise user privacy and depend heavily on network conditions.
  • Edge inference places models partially or fully on mobile devices to make predictions locally.

3.1 AI Computing Hardware in Edge Computing

Edge AI hardware is evolving to address bandwidth, privacy, and compute–transmission constraints through lightweight servers and specialized processors. Representative platforms span VPUs, Edge TPUs, GPU-equipped systems, mobile SoCs, and NPUs optimized for efficient neural computation.

  • Hardware motivations and categories: Edge AI hardware development responds to bandwidth, privacy, and compute–transmission challenges, using lightweight servers such as Raspberry Pi and NVIDIA TX2 alongside specialized hardware.The survey focuses on representative function-oriented hardware, while broader edge-computing servers are referenced separately.
  • Vision processing units: VPUs provide power-efficient execution of edge vision workloads; Intel’s Neural Compute Stick uses the Movidius Myriad X and supports Linux, Windows, Raspbian, Raspberry Pi, and Intel NUC systems.The device is plug-and-play and targets a balance between power-supply efficiency and computing performance.
  • Edge tensor processing units: Google’s Edge TPU accelerates edge inference for compressed CNNs, TensorFlow Lite models, and real-world detection and segmentation, but remains unavailable to the public.Supported models include MobileNets, MobileNets SSD, and Inception.
  • Neural processing units: NPUs improve edge neural efficiency through resource-aware configuration, data reuse, and skipped unnecessary computations to reduce memory-bandwidth demands.ARM Ethos N-777 delivers up to 4 TOPS and scales to hundreds of TOPS in multicore deployments,.

3.2 Efficient Network Architecture

Efficient network design spans manually crafted lightweight architectures and automated neural architecture search (NAS). Representative methods compress or factorize convolutions, while NAS generates and evaluates subnetworks but faces challenges from non-differentiable objectives [126]–.

  • Efficient Network Architecture: Lightweight architectures reduce model complexity through squeeze–expansion layers in SqueezeNet and depth-wise plus point-wise convolution in MobileNet.MobileNet’s decomposition is motivated by the low-rank property of convolution kernels, making it approximately equivalent to conventional convolution.
  • Efficient Network Architecture: Neural architecture search (NAS) automatically explores efficient architectures by generating, training, and evaluating subnetworks with an RNN controller [123].A major NAS challenge is the non-differentiability of the subnetwork-performance objective [126]–.

3.3 Compression

Model compression reduces DNN complexity through knowledge distillation, quantization, and pruning. These approaches transfer knowledge, lower numerical precision, or remove computation paths while balancing efficiency and accuracy.

  • Knowledge Distillation: Knowledge distillation transfers knowledge from complex models or ensembles to lightweight models through response-, feature-, or relation-based distillation [134]–[147].Response-based KD uses network outputs as soft targets; feature-based KD matches intermediate feature maps, while relation-based KD captures relationships among layers or data samples.
  • Quantization: Quantization improves training and inference efficiency through half- and mixed-precision training, while parameter quantization uses QAT or PTQ [150]–.QAT quantizes parameters during training with floating-point backpropagation, whereas PTQ avoids fine-tuning and has negligible overhead but often lower accuracy, especially at low precision.
  • Pruning: Pruning reduces parameter space by removing computation paths through one-time or runtime pruning –.One-time methods prune after training, jointly learn sparse structures, or prune at initialization; runtime methods dynamically select paths for inference under limited computation budgets.

4 EDGE-CLOUD COLLABORATIVE AI

Edge-cloud collaborative AI spans privacy-primary federated learning and efficiency-primary collaboration, addressing heterogeneous data, communication limits, personalization, and deployment constraints. Emerging approaches include split deployment, personalized and bidirectional modeling, and transfer mechanisms for adapting models across cloud and edge environments.

  • Privacy-primary collaboration: Federated learning keeps client data local and aggregates models in the cloud, making edge-cloud collaboration practical for privacy-sensitive applications.It includes cross-device settings with phones, laptops, or IoT devices and cross-silo settings with organizational data silos.
  • Federated learning challenges: Non-IID edge data creates heterogeneous models and gradients through feature skew, label skew, and differing feature-label relationships.These distribution differences are a central challenge for collaborative learning across devices.
  • Federated learning challenges: Communication-efficient federated learning accelerates convergence or introduces intermediate servers for partial aggregation under limited communication budgets.Hierarchical federation is described as scalable in many practical cases.
  • Federated learning challenges: Federated learning remains vulnerable to poisoning attacks that threaten robustness and inference attacks that threaten privacy.Inference methods such as Deep Leakage from Gradients can retrieve raw images and texts, motivating encryption and secure multiparty computation [218], –.
  • Efficiency-primary collaboration: Backbone-patch decomposition with MetaPatch and MoModistill reduces edge computation, preserves personalization, and outperforms baselines on recommendation benchmarks.The framework uses a backbone-patch decomposition, re-calibrating the backbone to avoid local optima.
  • Collaborative learning perspectives: Heterogeneous transfer learning, meta-learning, and causal inference support bidirectional adaptation across differing edge and cloud data distributions.These perspectives target domain-invariant representations, rapid adaptation to new edge tasks, bias reduction, fairness, and robustness to confounding effects –, [256],.

5 ADVANCED TOPICS

Advanced edge AI topics center on adapting foundation models to constrained devices, distributing and compressing graph neural networks, and extending reinforcement learning through federated edge-cloud collaboration. These directions address limited storage, bandwidth, computation, data, and samples while preserving collaborative learning capabilities.

  • Pretrained Models: Foundation models such as BERT and GPT-3 offer versatile few-shot and zero-shot transfer, but their growing scale creates edge deployment constraints.Larger models trained on larger corpora generally achieve better downstream performance, motivating efficient edge adaptation.
  • Pretrained Models: Compressing foundation models is necessary for edge mobiles with limited storage or network bandwidth, using quantization, pruning, and Transformer-specific parameter sharing.Reducing hidden sizes or layers can also accelerate inference, while PABEE dynamically exits once intermediate predictions remain unchanged for several steps.
  • Pretrained Models: Parameter-efficient adaptation is important because full fine-tuning wastes storage and bandwidth and is poorly sample-efficient in few-sample edge environments [285].Traditional fine-tuning updates almost all foundation-model parameters and creates an independently stored model for each downstream task.
  • Graph Neural Networks: Federated GNNs keep graph data on edges while aggregating local updates into a shared global model, with methods addressing distributed security, privacy, heterogeneity, and decentralized multi-task learning [296]–[309].Examples include Feddy [296], AS-FGNN, FedGNN [300], clustering methods FedCG and GCFL [308], global self-supervision in FedGL [304], and serverless SpreadGNN [305].
  • Graph Neural Networks: GNN-specific quantization and meta learning target edge memory and limited samples, with Bi-GCN reducing memory consumption 30x and accelerating inference about 47x.SGQuant uses component-wise, topology-aware, and layer-wise quantization, while existing GNN meta-learning methods –[321] largely adopt Model-Agnostic Meta-Learning [257].
  • Reinforcement Learning: Federated reinforcement learning extends edge-cloud collaboration through horizontal and vertical settings, supporting distributed agents with similar tasks or environments forming a global system –[330].HFRL addresses non-i.i.d. data and distributed optimization, while VFRL models partially observable environments; other approaches share reward shaping or policy gradients, [332].

6 FUTURE DIRECTIONS AND CONCLUSION

The section identifies scarce open-source datasets and immature software platforms as key barriers to Edge-Cloud collaborative AI, while highlighting opportunities and challenges across recommendation, autonomous driving, gaming, and IoT security. It concludes by framing cloud AI, edge AI, and collaborative AI as three paradigms reviewed across computer vision, NLP, and web services.

  • Challenges: Scarce open-source datasets limit academic exploration of Edge-Cloud collaborative AI because edge-side fine-grained data are often not transmitted due to communication costs and serving constraints.The survey notes that only one dataset is available for Mobile Edge Intelligence in recommendation.
  • Challenges: Well-established software platforms are still lacking, making it costly to construct collaboration environments and simulate heterogeneous edges, communication noise, and real-world federated-learning behavior.Federated learning must handle uncontrollable numbers of local models uploaded from edge devices, which may affect convergence.
  • Applications: Edge-Cloud collaboration offers application benefits while addressing persistent issues in recommendation systems, autonomous driving, and games, including bias, privacy leakage, latency, and difficult cloud–edge work division.The supplied passages describe recommendation, autonomous-driving, and gaming scenarios as application areas for collaboration.
  • Security: Collaborative learning can improve IoT security by avoiding direct transfer of sensitive data and enabling trusted cloud models to identify malicious edges, although cloud-side detection may be computationally intensive for edge devices.The mechanism transfers model parameters, latent representations, and gradients rather than sensitive data.
  • Conclusion: The survey organizes AI into cloud, edge, and Edge-Cloud collaborative paradigms, systematically reviewing their advances across computer vision, natural language processing, and cloud-powered web services.It presents this scope as a comprehensive account of the underlying polarization and collaboration among the paradigms.
Loading 2111.06061v3…