CoChat
Sign Up
Cited by 2,455IoT and Edge/Fog ComputingAdvanced Neural Network ApplicationsInternet of Things and AI

Edge Intelligence: Paving the Last Mile of Artificial Intelligence With Edge Computing

Open access
Zhou, Zhi, Chen, Xu, Li, En, Zeng, Liekang, Luo, Ke, Zhang, Junshan·2019·Proceedings of the IEEE
DOI 10.1109/jproc.2019.2918951
Open this paper in CoChat — free accountOpen PDF

Abstract

With the breakthroughs in deep learning, the recent years have witnessed a booming of artificial intelligence (AI) applications and services, spanning from personal assistant to recommendation systems to video/audio surveillance. More recently, with the proliferation of mobile computing and Internet of Things (IoT), billions of mobile and IoT devices are connected to the Internet, generating zillions bytes of data at the network edge. Driving by this trend, there is an urgent need to push the AI frontiers to the network edge so as to fully unleash the potential of the edge big data. To meet this demand, edge computing, an emerging paradigm that pushes computing tasks and services from the network core to the network edge, has been widely recognized as a promising solution. The resulted new interdiscipline, edge AI or edge intelligence (EI), is beginning to receive a tremendous amount of interest. However, research on EI is still in its infancy stage, and a dedicated venue for exchanging the recent advances of EI is highly desired by both the computer system and AI communities. To this end, we conduct a comprehensive survey of the recent research efforts on EI. Specifically, we first review the background and motivation for AI running at the network edge. We then provide an overview of the overarching architectures, frameworks, and emerging key technologies for deep learning model toward training/inference at the network edge. Finally, we discuss future research opportunities on EI. We believe that this survey will elicit escalating attentions, stimulate fruitful discussions, and inspire further research ideas on EI.

Integrity signals

Registry

Crossref · 10.1109/jproc.2019.2918951

Cites

139 of 139 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

10 of 54 cited sources checked · 4 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Stamoulis et al. achieve an improvement by up to 6× in terms of minimum energy per image under accuracy constraints by casting adaptive DNN model selection as a hyperparameter optimization problem and using Bayesian optimization.

    “Stamoulis et al. [106] cast the adaptive DNN model selection issue as a hyperparameter optimization problem by taking into account the accuracy and communication constraints imposed by the devices.Then, a Bayesian optimization (BO) is adopted to solve this problem, achieving an improvement by up to 6× in terms of minimum energy per image under accuracy constraints.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [106] 106. D. Stamoulis et al., "Designing adaptive neural networks for energy-constrained image classification," in Proc. ACM ICCAD, 2018, Art. no. 23. source
  • Premisecitation

    The performance of transfer learning varies with architectures and transfer techniques.

    “The result varies with architectures and transfer techniques.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [62] 62. R. Sharma, S. Biookaghazadeh, B. Li, and M. Zhao, "Are existing knowledge transfer techniques effective for deep learning with edge devices?" in Proc. IEEE Int. Conf. Edge Comput. (EDGE), Jul. 2018, pp. 42-49. source
    • cites [44] 44. Q. Chen, Z. Zheng, C. Hu, D. Wang, and F. Liu, "Data-driven task allocation for multi-task transfer learning on the edge," in Proc. IEEE 39th Int. Conf. Distrib. Comput. Syst. (ICDCS), 2019. source

      Section 5.2.2 (p. 6) reports accuracy for one architecture under different distributed-training conditions, but does not compare architectures or transfer-learning techniques. Section 6 (p. 6) calls studies of other DNN architectures future work. Searches for transfer learning and variation across architectures or transfer techniques found no passage addressing the claim. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • Premisecitation

    Cachier can increase responsiveness by 3× or more.

    “Cachier can increase responsiveness by 3× or more.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [93] 93. U. Drolia, K. Guo, J. Tan, R. Gandhi, and P. Narasimhan, "Cachier: Edge-caching for recognition applications," in Proc. IEEE ICDCS, Jun. 2017, pp. 276-286. source

      The Results section discusses Edgent’s inference latency, but does not report a result for Cachier. Searches for “Cachier” and “responsiveness three times” found no passage supporting a 3× improvement. (LLM, conf=0.77)

      Experimental — accuracy not yet measured
  • Interpretationreasoning

    Data generated at the network edge need AI to fully unlock their potential.

    “Data Generated at the Network Edge Need AI to Fully Unlock Their Potential”
    • Also stated in: § A. Motivation and Benefits of Edge Intelligence · boosted
      “AI will be functionally necessary due to its ability to quickly analyze those huge data volumes and extract insights from them for high-quality decision making.”
    Quote in the paperSubstantial claim
  • Premisecitation

    Cisco estimates that nearly 850 ZB will be generated by all people, machines, and things at the network edge by 2021.

    “Cisco estimates that nearly 850 ZB will be generated by all people, machines, and things at the network edge by 2021 [3].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] 3. Cisco Global Cloud Index: Forecast and Methodology, 2016-2021, White Paper. [Online]. Available: https://www.cisco.com/c/en/us/ solutions/collateral/service-provider/global- cloud-index-gci/white-paper-c11-738085.html

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. 28
  • May combine two claims. 27

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    VGG-16 requires 138 million weights and 15.5G MACs to classify an image of size 224 × 224.

    “VGG-16 [19] is trained to a deeper structure of 16 layers consisting of 13 convolution layers and three fully connected layers, requiring 138 million weights and 15.5G MACs to classify the image with a size of 224 × 224.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [19] 19. K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," 2014, arXiv:1409.1556. [Online]. Available: https://arxiv.org/abs/1409.1556 source

      Table 1 (p. 3) identifies configuration D as a 16-weight-layer network with a 224 × 224 RGB input, and Table 2 (p. 3) lists 138 million parameters for D. Section 2.1 (p. 2) describes the three fully connected layers. Our copy does not report the claimed 15.5G MACs per image, so that part is unsupported. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    FedCS allows the centralized server to aggregate as many client updates as possible and to accelerate performance improvement in ML models.

    “In particular, an update aggregation protocol named FedCS is developed to allow the centralized server to aggregate as many client updates as possible and to accelerate performance improvement in ML models.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [52] 52. T. Nishio and R. Yonetani, "Client selection for federated learning with heterogeneous resources in mobile edge," 2018, arXiv:1804.08333. [Online]. Available: https://arxiv.org/abs/ 1804.08333 source

      The Introduction (p. 2) describes pruning, quantization, and Huffman coding, while §6.3 (pp. 9–10) discusses inference speedups from compression. Searches for FedCS, client-update aggregation, and federated client selection found no passage addressing the claimed protocol or its effect on model improvement. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    The optimization in [45] emphasizes the properties of unbalanced and non-IID since the distributed data may come from various sources.

    “The optimization in [45] emphasizes the properties of unbalanced and non-IID since the distributed data may come from various sources.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [45] 45. H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. Y. Arcas, "Communication-efficient learning of deep networks from decentralized data," 2016, arXiv:1602.05629. [Online]. Available: https://arxiv.org/abs/1602.05629 source

  • May combine two claims.

    Premisecitation

    Moving a tremendous amount of data across the wide area network can incur prohibitively high monetary cost and transmission delay, and privacy leakage can be a major concern.

    “when moving a tremendous amount of data across the wide area network (WAN), both monetary cost and transmission delay can be prohibitively high, and the privacy leakage can also be a major concern [5].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] 5. Q. Pu et al., "Low latency geo-distributed data analytics," in Proc. ACM SIGCOMM, 2015, pp. 421-434. source

      The Introduction and Section 6 (Advantages and disadvantages; page numbers unavailable) discuss generative models and their computational properties, not transferring large datasets over a WAN. Searches for WAN costs, transmission delay, and privacy leakage found no passage addressing the claim. (LLM, conf=0.75)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Physical proximity between computing and information-generation sources promises benefits compared with traditional cloud-based computing, including low latency, energy efficiency, privacy protection, reduced bandwidth consumption, on-premises operation, and context awareness.

    “the physical proximity between the computing and information-generation sources promises several benefits compared to the traditional cloud-based computing paradigm, including low latency, energy efficiency, privacy protection, reduced bandwidth consumption, on-premises, and context awareness [6], [8].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [6] 6. W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, "Edge computing: Vision and challenges," IEEE Internet Things J., vol. 3, no. 5, pp. 637-646, Oct. 2016. source
    • cites [8] 8. Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, "A survey on mobile edge computing: The communication perspective," IEEE Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322-2358, 4th Quart., 2017. source

  • May combine two claims.

    Premisecitation

    A good performance improvement is obtained by transferring knowledge from both the intermediate layers and last layer of the teacher to a shallower student, while other architectures and transfer techniques do not fare so well and some even lead to negative performance impact.

    “A good performance improvement is obtained by transferring knowledge from both the intermediate layers and last layer of the teacher to a shallower student while other architectures and transfer techniques do not fare so well and some of them even lead to negative performance impact.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [62] 62. R. Sharma, S. Biookaghazadeh, B. Li, and M. Zhao, "Are existing knowledge transfer techniques effective for deep learning with edge devices?" in Proc. IEEE Int. Conf. Edge Comput. (EDGE), Jul. 2018, pp. 42-49. source
    • cites [44] 44. Q. Chen, Z. Zheng, C. Hu, D. Wang, and F. Liu, "Data-driven task allocation for multi-task transfer learning on the edge," in Proc. IEEE 39th Int. Conf. Distrib. Comput. Syst. (ICDCS), 2019. source

      Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Gossip averaging can fast converge toward a consensus among nodes by exchanging information peer to peer.

    “The early work on random gossip algorithms is gossip averaging [63], which can fast converge toward a consensus among nodes by exchanging information peer to peer.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [63] 63. S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2508-2530, Jun. 2006. source

  • May combine two claims.

    Premisecitation

    The control algorithm proposed by Wang et al. [51] is based on convergence analysis of distributed gradient descent and can be applied to federated learning in edge computing with provable convergence.

    “The algorithm is based on the convergence analysis of distributed gradient descent and can be applied to federated learning in edge computing with provable convergence.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [51] 51. S. Wang et al., "Adaptive federated learning in resource constrained edge computing systems," IEEE J. Sel. Areas Commun., vol. 37, no. 3, pp. 1205-1221, Jun. 2019.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    As the level of EI goes higher, the transmission latency of data offloading decreases, data privacy increases, and WAN bandwidth cost reduces.

    “As a result, the transmission latency of data offloading decreases, the data privacy increases, and the WAN bandwidth cost reduces.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The main challenges of federated learning are optimization and communication.

    “The main challenges of federated learning are optimization and communication.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Edge computing and AI complement each other from a technical perspective, and their application and popularization are mutually beneficial.

    “While edge computing and AI complement each other from a technical perspective, their application and popularization are also mutually beneficial.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Federated learning is built upon a central server for aggregating local updates.

    “Though the federated learning technique exploits a new decentralized deep learning architecture, it is built upon a central server for aggregating local updates.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Gartner forecasts that more than 80% of enterprise IoT projects will include an AI component by 2022, up from only 10% today.

    “As forecast by Gartner [33], more than 80% of enterprise IoT projects will include an AI component by 2022, up from only 10% today.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] 33. AI Trends for Enterprise Computing. [Online]. Available: https://www.gartner.com/ smarterwithgartner/3-ai-trends-for-enterprise- computing/

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    There is no single best level of EI in general; the best level is application-dependent and should be determined by jointly considering multiple criteria such as latency, energy efficiency, privacy, and WAN bandwidth cost.

    “This conflict indicates that there is no”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Energy efficiency is mainly affected by the size of the target training model and resources of the used devices.

    “Energy efficiency is mainly affected by the size of the target training model and resources of the used devices.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ResNet reaches human-level accuracy with a top-5 error rate below 5%.

    “ResNet [20], the state-of-the-art effort, uses the”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [20] 20. K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE CVPR, Jun. 2016, pp. 770-778. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Real-time video analytics is envisioned to be a killer application for edge computing.

    “real-time video analytics is envisioned to be a killer application for edge computing [12], [37], [38].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [12] 12. G. Ananthanarayanan et al., "Real-time video analytics: The killer app for edge computing," Computer, vol. 50, no. 10, pp. 58-67, 2017. source
    • cites [37] 37. H. Zhang, G. Ananthanarayanan, P. Bodik, M. Philipose, P. Bahl, and M. J. Freedman, "Live video analytics at scale with approximation and delay-tolerance," in Proc. USENIX NSDI, 2017, pp. 377-392.
    • cites [38] 38. C.-C. Hung et al., "VideoEdge: Processing camera streams using hierarchical clusters," in Proc. IEEE/ACM Symp. Edge Comput. (SEC), Oct. 2018, pp. 115-131. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AlexNet requires 61 million weights and 724 million MACs to classify an image of size 227 × 227.

    “AlexNet requires 61 million weights and 724 million MACs (multiply-add computation) to classify the image with a size of 227 × 227.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] 18. A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," in Proc. NIPS, 2012, pp. 1097-1105.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    Gaia focuses on geodistributed datacenters that are capacity-unlimited, making it not generally applicable to edge computing nodes whose capacity is highly constrained.

    “However, Gaia focuses on geodistributed datacenters that are capacity-unlimited, making it is not generally applicable to edge computing nodes whose capacity is highly constrained.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] 50. K. Hsieh, A. Harlap, N. Vijaykumar, D. Konomis, G. R. Ganger, and P. B. Gibbons, "Gaia: Geo-distributed machine learning approaching LAN speeds," in Proc. NSDI, 2017, pp. 629-647.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Under decentralized training mode, the convergence value depends on the way the gradient is synchronized and updated.

    “Under the decentralized training mode, the convergence value depends on the way the gradient is synchronized and updated.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    When training a DNN model using data originated at massive end devices, raw or intermediate data should be transferred out of the end devices.

    “When training the DNN model by using the data originated at a massive of end devices, the raw data or intermediate data should be transferred out of the end devices.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Deep learning represents state-of-the-art AI technology and is a highly resource-demanding workload that naturally suits edge computing.

    “Deep learning represents the state-of-the-art AI technology as well as a highly resource-demanding workload that naturally suits for edge computing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    When training a DNN model in a decentralized manner, both computation and communication processes consume enormous energy.

    “When training the DNN model in a decentralized manner, both the computation and communication processes consume enormous energy.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: For most end devices, they are energy-constrained.

  • May combine two claims.

    Interpretationreasoning

    As the level of EI goes higher, computational latency and energy consumption increase.

    “However, this is achieved at the cost of increased computational latency and energy consumption.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    GoogleNet achieves better accuracy than VGG-16 while requiring seven million weights and 1.43G MACs to process an image of the same size.

    “GoogleNet achieves a better accuracy performance than VGG-16, while only requiring seven million weights and 1.43G MACs to process the image with the same size.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28] 28. C. Szegedy et al., "Going deeper with convolutions," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2015, pp. 1-9. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    MeDNN improves DNN model inference by 1.86-2.44× with two to four worker nodes and saves 26.5% of additional computing time and 14.2% of extra communication time.

    “MeDNN improves the DNN model inference by 1.86-2.44×with two to four worker nodes and saves 26.5% of additional computing time and 14.2% of extra communication time.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [84] 84. J. Mao et al., "MeDNN: A distributed mobile system with enhanced partition and deployment for large-scale DNNs," in Proc. 36th Int. Conf. Comput.-Aided Design, Nov. 2017, pp. 751-756. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Communication overhead is affected by the size of the original input data, the way of transmission, and the available bandwidth.

    “Communication overhead is affected by the size of the original input data, the way of transmission, and the available bandwidth.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    The high-dimensional configuration problem involves a huge search space of parameters and is very challenging to tackle due to its combinatorial nature.

    “Due to the combinatorial nature, the high-dimensional configuration problem involves a huge search space of parameters and is very challenging to tackle.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Standard performance indicators such as top-k accuracy or mean average precision fail to reflect the runtime performance of DNN model inference on edge devices.

  • May combine two claims.

    Premisenone given

    Data parallelism may bring heavy overhead of communication, while model parallelism usually leads to severe underutilization of computation resources.

    “However, data parallelism may bring heavy overhead of communication, while model parallelism usually leads to severe underutilization of computation resources.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    PipeDream shows the advantage of reducing communication overhead and utilizing computing resource efficiently.

    “PipeDream shows the advantage of reducing communication overhead and utilizing computing resource efficiently.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [61] 61. A. Harlap, D. Narayanan, A. Phanishayee, V. Seshadri, G. R. Ganger, and P. B. Gibbons, "PipeDream: Fast and efficient pipeline parallel DNN training," 2018, arXiv:1806.03377. [Online]. Available: https://arxiv.org/abs/1806.03377 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Communication latency may vary with the size of transmitted raw or intermediate data and the bandwidth of the network connection.

    “The communication latency may vary from the size of transmitted raw or intermediate data and the bandwidth of network connection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    By deployment on large-scale systems, trivial gossip-based algorithms at scale lead to a communication imbalance, poor convergence, and heavy communication overhead.

    “By deployment on large-scale systems, Daily et al. [66] show that the trivial gossip-based algorithms at scale lead to a communication imbalance, poor convergence, and heavy communication overhead.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [66] 66. J. Daily, A. Vishnu, C. Siegel, T. Warfel, and V. Amatya, "GossipGraD: Scalable deep learning using gossip communication based asynchronous gradient descent," 2018, arXiv:1803.05880. [Online]. Available: https://arxiv.org/ abs/1803.05880 source

  • May combine two claims.

    Premisecitation

    FoggyCache reduces computation latency and energy consumption by a factor of 3× to 10×.

    “FoggyCache reduces computation latency and energy consumption by a factor of 3× to 10×.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [95] 95. P. Guo, B. Hu, R. Li, and W. Hu, "FoggyCache: Cross-device approximate computation reuse," in Proc. ACM Mobicom, 2018, pp. 19-34. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Latency is one of the most fundamental performance indicators of distributed DNN model training because it directly influences when the trained model is available for use.

    “Arguably, latency is one of the most fundamental performance indicators of distributed DNN model training since it directly influences when the trained model is available for use.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    A DNN model can be split inside between two successive layers with two partitions deployed on different locations without losing accuracy.

    “This is based on the important observation that a DNN model can be split inside between two successive layers with two partitions deployed on different locations without losing accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    AdaDeep formulates goals and constraints on accuracy, latency, storage, and energy into a unified optimization problem and uses DRL to find a good combination of compression techniques.

    “the proposed automatic optimization framework AdaDeep [75] systematically formulates the goals and constraints on accuracy, latency, storage, and energy into a unified optimization problem and leverages DRL to effectively find a good combination of compression techniques.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [75] 75. S. Liu, Y. Lin, Z. Zhou, K. Nan, H. Liu, and J. Du, "On-demand deep model compression for mobile devices: A usage-driven model selection framework," in Proc. 16th Annu. Int. Conf. Mobile Syst., Appl., Services, 2018, pp. 389-400. source

  • May combine two claims.

    Interpretationreasoning

    The decentralized architecture is naturally friendly to user's privacy, for which systems based on the decentralized architecture such as BlockFL and GossipGraD typically preserve privacy better.

    “The decentralized architecture is naturally friendly to user's privacy, for which the systems that are based on the decentralized architecture such as BlockFL and GossipGraD typically preserve privacy better.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Communication overhead increases training latency, energy, and bandwidth consumption.

    “Intuitively, this communication overhead increases the training latency, energy, and bandwidth consumption.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Precog reaches a speedup of 5× by using predictions of Markov chains to prefetch data onto the mobile device.

    “Precog uses predictions of Markov chains to prefetch data onto the mobile device and reach a speedup of 5×.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94] 94. U. Drolia, K. Guo, and P . Narasimhan, "Precog: Prefetching for image recognition applications at the edge," in Proc. ACM/IEEE Symp. Edge Comput. (SEC), Oct. 2017, p. 17. source

  • May combine two claims.

    Interpretationreasoning

    Compared with DNN training under the cloud-based framework, DNN training under edge-based framework pays more attention to protecting users' privacy and training an available deep learning model faster.

    “Compared with the DNN training under the cloud-based framework, the DNN training under edge-based framework pays more attention to protecting users' privacy and training an available deep learning model faster.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    The latency of distributed training typically consists of computation latency and communication latency.

    “The latency of the distributed training process typically consists of the computation latency and the communication latency.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    Both Blot et al. [64] and Jin et al. [65] apply gossip algorithms on the updates of SGD, but neither of them has performance convergence degradation at large scale.

    “Both Blot et al. [64] and Jin et al. [65] apply gossip algorithms on the updates of SGD, but neither of them has performance convergence degradation at large scale.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [64] 64. M. Blot, D. Picard, M. Cord, and N. Thome, "Gossip training for deep learning," 2016, arXiv:1611.09726. [Online]. Available: https://arxiv.org/abs/1611.09726 source
    • cites [65] 65. P. H. Jin, Q. Yuan, F. Iandola, and K. Keutzer, "How to scale distributed deep learning?" 2016, arXiv:1611.04581. [Online]. Available: https://arxiv.org/abs/1611.04581 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ReXCam reduces computation workload by 4.6× and improves DNN model inference accuracy by 27%.

    “ReXCam reduces computation workload by 4.6× and improves DNN model inference accuracy by 27%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] 99. S. Jain, J. Jiang, Y. Shu, G. Ananthanarayanan, and J. Gonzalez, "ReXCam: Resource-efficient, cross-camera video analytics at enterprise scale," 2018, arXiv:1811.01268. [Online]. Available: https://arxiv.org/abs/1811.01268 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Huang et al. found a new combination of factors (number of proposals, input image size, and feature extractor) that outperforms the state-of-the-art method for object recognition.

    “For the object recognition application, Huang et al. [120] investigate the influence of the main factors, e.g., number of proposals, input image size, and the selection of feature extractor, on inference speed and accuracy.Based on their experiment results, a new combination of these factors is found to outperform the state-of-the-art method.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [120] 120. J. Huang et al., "Speed/accuracy trade-offs for modern convolutional object detectors," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 7310-7311. source
    • Rests on an assumption: Standard performance indicators such as top-k accuracy or mean average precision fail to reflect the runtime performance of DNN model inference on edge devices.

  • May combine two claims.

    Premisecitation

    IONN significantly improves query performance and energy consumption during DNN model uploading compared to uploading the entire model.

    “Compared to the approach that uploads the entire model, IONN significantly improves query performance and energy consumption during DNN model uploading.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] 82. H.-J. Jeong, H.-J. Lee, C. H. Shin, and S.-M. Moon, "IONN: Incremental offloading of neural network computations from mobile devices to edge servers," in Proc. ACM Symp. Cloud Comput., 2018, pp. 401-411. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Gossip distributed algorithms enjoy the advantage of full asynchronization and total decentralization as they have no requirement on centralized nodes or variables.

    “The gossip distributed algorithms enjoy the advantage of full asynchronization and total decentralization as they have no requirement on centralized nodes or variables.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The cascading network achieves a speedup of 20% by adding the MP layer and fully connected layer to the standard DNN model.

    “the cascading network [91] simply adds the MP layer and fully connected layer to the standard DNN model and achieves a speedup of 20%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [91] 91. S. Leroux et al., "The cascading neural network: Building the Internet of smart things," Knowl. Inf. Syst., vol. 52, no. 3, pp. 791-814, 2017.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Transfer learning tends to work if the features are general (suitable to both base and target tasks) instead of specific to the base task.

    “This process will tend to work if the features are general (i.e., suitable to both base and target tasks) instead of specific to the base task.”
    Quote in the paperNot a substantial claim
48 of 48

  • The paper relies on this without support.

    There is a lack of a dedicated venue for summarizing, discussing, and disseminating recent advances of edge intelligence in both industry and academia.

    “There is, in general, a lack of venue dedicated for summarizing, discussing, and disseminating the recent advances of EI, in both industrial and academia.”

  • The paper relies on this without support.

    Research and practice on edge intelligence are still in a very early stage.

    “Notably, research and practice on this emerging interdiscipline-EI-are still in a very early stage.”

  • The paper relies on this without support.

    The edge ecosystem will present many novel application scenarios for AI and fuel the continuous booming of AI.

    “Clearly, via bringing the huge volumes of data to AI, the edge ecosystem will present many novel application scenarios for AI and fuel the continuous booming of AI.”

  • The paper relies on this without support.

    For most end devices, they are energy-constrained.

    “However, for most end devices, they are energy-constrained.”

  • The paper relies on this without support.

    It is inevitable to deal with privacy issues when raw or intermediate data are transferred out of end devices.

    “Obviously, it is inevitable to deal with privacy issues in this scenario.”

  • The paper relies on this without support.

    For some real-time intelligent mobile applications, they usually have stringent deadline requirement such as 100-ms latency.

    “For some real-time intelligent mobile applications (e.g., AR/VR mobile gaming and intelligent robots), they usually have stringent deadline requirement such as 100-ms latency.”

  • The paper relies on this without support.

    For some mobile applications requiring high reliability, they demand ultrahigh accuracy on DNN model inference.

    “For some mobile applications requiring a high level of reliability, such as self-driving car and face authentication, they demand the ultrahigh accuracy on the DNN model inference.”

  • The paper relies on this without support.

    Most existing efforts use a single compression technique.

    “While most existing efforts use a single compression technique, they may not suffice to meet the diverse requirements and constraints on accuracy, latency, storage, and energy imposed by some IoT devices.”

  • The paper relies on this without support.

    Standard performance indicators such as top-k accuracy or mean average precision fail to reflect the runtime performance of DNN model inference on edge devices.

    “However, it is difficult for software developers to choose an appropriate DNN model for the EI application because the standard performance indicators such as top-k accuracy or mean average precision fail to reflect the runtime performance of DNN model inference on edge devices.”

  • The paper relies on this without support.

    Most current edge AI programming/software platforms mainly serve as relays for connecting to cloud data centers.

    “Nevertheless, currently, most of these platforms mainly serve as relays for connecting to the powerful cloud data centers.”

  • The paper relies on this without support.

    Most deep learning-based AI models are highly resource-intensive.

    “Most of the deep learning-based AI models are highly resourceintensive, which means that powerful computing capability supported by abundant hardware resources (e.g., GPU, field-programmable gate array (FPGA), and TPU) is important to boost the performance of these AI models.”

  • The paper relies on this without support.

    For an EI application with a specific mission, there is usually a series of DNN model candidates that are capable of finishing the task.

    “For an EI application with a specific mission, there is usually a series of DNN model candidates that are capable of finishing the task.”

  • The paper relies on this without support.

    The service environments for many EI application scenarios are of high dynamics and it is hard to accurately predict future events.

    “Since for many EI application scenarios (e.g., smart cities), the service environments are of high dynamics and it is hard to accurately predict future events.”

  • The paper relies on this without support.

    The high-efficiency operation of EI services may require close collaboration and integration across different service providers.

    “The high-efficiency operation of EI services may require close collaboration and integration across different service providers, e.g., for implementing expanded resource sharing and smooth service handover.”

  • The paper relies on this without support.

    A user can be a service consumer and, meanwhile, a data generator as well.

    “Also, for EI service, a user can be a service consumer and, meanwhile, a data generator as well.”

  • The paper relies on this without support.

    There is a stringent need to push the AI frontier from the cloud to the network edge.

    “Driving by flourishing of both AI and IoT, there is a stringent need to pushing the AI frontier from the cloud to the network edge.”

  • The paper relies on this without support.

    Edge computing is widely recognized as a promising solution to support computation-intensive AI applications in resource-constrained environments.

    “To fulfill this trend, edge computing has been widely recognized as a promising solution to support computation-intensive AI applications in resource-constrained environments.”
17 of 17

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • UAVs generate lots of data at the network edge, and directly sharing original data sets between many edge nodes has a high risk associated with it.

    “Additionally, UAVs generate lots of data at the network edge and directly sharing original data sets between many edge nodes has a high risk associated with it [8].”

    From A Survey on the Convergence of Edge Computing and AI for UAVs: Opportunities and Challenges (2022) · cites this paper as [8]

    The Security and Privacy Issues section says end users and devices generate massive amounts of data at the network edge and that sharing original data sets among multiple edge nodes risks privacy leakage. It does not make that combined claim specifically about UAVs; its discussion of drones in the Input Filtering section concerns video transmission and bandwidth, not data-sharing privacy risk. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Edge Intelligence (EI) proposes a new paradigm combining AI and edge computing, with the goal of performing distributed computing of DL models.

    “Edge Intelligence (EI) Zhou et al. (2019) proposes a new paradigm combining AI and edge computing, with the goal of performing distributed computing of DL models.”

    From Deepfakes: current and future trends (2024) · cites this paper as Zhou et al. 2019

  • Edge computing entails advantages such as low latency, energy and bandwidth efficiency, privacy protection, and context awareness.

    “This entails some advantages such as low latency, energy and bandwidth efficiency, privacy protection, and context awareness (Zhou et al. 2019).”

    From Deepfakes: current and future trends (2024) · cites this paper as Zhou et al. 2019

  • In edge computing environments where distributed models are trained locally and updates are aggregated centrally, control of updates aggregation frequency has a significant influence on communication overhead.

    “In the case of training a deep learning model in an edge computing environment, where distributed models are trained locally first and then updates are aggregated centrally (e.g., federated learning), the control of updates aggregation frequency has a significant influence on the communication overhead.”

    From A Survey on the Convergence of Edge Computing and AI for UAVs: Opportunities and Challenges (2022) · cites this paper as [8]

  • Edge AI provides a promising solution for connected intelligence by enabling data collection, processing, transmission, and consumption at the network edge.

    “Edge AI provides a promising solution for connected intelligence by enabling data collection, processing, transmission, and consumption at the network edge [29], [30].”

    From Edge artificial intelligence for 6g: vision, enabling technologies, and applications (2022) · cites this paper as [30]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Inference metrics are counted as five but six are listed

    Minor

    The section introduces five inference metrics, but its numbered list runs from latency through memory footprint and contains six.

1 of 1

Cites

Flagged first, then by citations

29 of the 168 references couldn’t be matched to a known paper.

Research works printed but not linked: 11

  1. 0, 3 AI Trends for Enterprise Computing
  2. 0, Microsoft Interactive Cloud Gaming
  3. 0, Democratizing AI
  4. hsieh, 2017, Gaia: Geo-distributed machine learning approaching LAN speeds, Proc NSDI, 629
  5. 2018, 5 Trends Emerge in the Gartner Hype Cycle for Emerging Technologies
  6. 0, Cisco global cloud index Forecast and methodology 2016–2021
  7. 0, Accelerating AI on the Intelligent Edge
  8. 2017, IEC White Paper Edge Intelligence
  9. 0, Edge Intelligence for Industrial Internet of Things
  10. han, 2015, Learning both weights and connections for efficient neural network, Proc Adv Neural Inf Process Syst, 1135
  11. 2018, 5 Trends Emerge in the Gartner Hype Cycle for Emerging Technologies

Cited by

Sorted by citations

Related

Authors
Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J.
Affiliations
  • Sun Yat-sen University (CN)
  • Arizona State University (US)
ORCID
Venue
Proceedings of the IEEE
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2019-08
Volume
107
Issue
8
Pages
1738-1762
ISSN
  • 0018-9219 (ISSN-L)
  • 1558-2256
DOI
10.1109/jproc.2019.2918951
OpenAlex
W2950865323
Semantic Scholar
928cd808aba140ec298508df87c5579811ff2f41
Cited by
2,455
References
168
Citations by year
2026: 455 · 2025: 352 · 2024: 332 · 2023: 363 · 2022: 370 · 2021: 341 · 2020: 215 · 2019: 26 · 2012: 1
FWCI
130.23
Type
journal article
Language
en
Open access
Licence
ieeexplore.ieee.org/Xplorehelp/downloads/license-information/OAPA.html (vor)
Locations
  • Proceedings of the IEEE (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Basic Research Program of China (2017YFB1001703)
  • National Natural Science Foundation of China (U1711265, 61802449)
  • Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2017ZT07X355)
  • Guangdong Natural Science Funds (2018A030313032)
  • Fundamental Research Funds for the Central Universities (17lgjc40)
  • Army Research Office (W911NF-16-1-0448)
  • Defense Threat Reduction Agency (HDTRA1-13-1-0029)
Topics
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Information Systems › Internet of Things and AI
Keywords
edge AI, edge computing, edge computing architecture, deep learning training, deep learning inference, Internet of Things, mobile computing

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review