Convergence of Edge Computing and Deep Learning: A Comprehensive Survey

In the paper

“As a typical and more widely used new form of applications [9], various deep learning-based intelligent services and applications have changed many aspects of people's lives due to the great advantages of Deep Learning (DL) in the fields of Computer Vision (CV) and Natural Language Processing (NLP) [10].”

What the sources say

• Partly supported

  The Introduction (p. 1) describes potential applications of fast object detection in driving, assistive devices, and robotics, supporting the computer-vision portion of the claim. It does not address natural language processing or establish that deep-learning applications have already changed many aspects of people's lives. (LLM, conf=0.86)

  cites [9] sourceExperimental — accuracy not yet measured
• Partly supported

  The Introduction (p. 4) reports deep learning outperforming other methods in numerous applications, and §5.22 (p. 25) describes advances in speech recognition and machine translation. Section 5.21 (p. 24) describes computer-vision applications in medical diagnosis, but frames their wider benefits as possibilities. These passages support the claimed advantages in vision and language-related tasks, not the assertion that applications have already changed many aspects of people’s lives. (LLM, conf=0.84)

  cites [10] sourceExperimental — accuracy not yet measured

In the paper

“With network slicing, customized and optimized resources can be provided for Edge DL services, which can help reduce latency caused by access networks and support dense access to these services [77].”

What the sources say

• ?Could not verify

  The Introduction (p. 2) discusses compressing neural networks for mobile devices, and Future Work (p. 12) discusses customized hardware for those compressed models. Neither addresses network slicing, access-network latency, or dense access to Edge DL services; searches for those terms and close paraphrases found no relevant passage. (LLM, conf=0.78)

  cites [77] sourceExperimental — accuracy not yet measured

In the paper

“For a computation-intensive task, it will be reasonably segmented and dispatched separately to the end, edge and cloud for execution, reducing the execution delay of the task while ensuring the accuracy of the results [12], [49], [50].”

What the sources say

• ?Not checked — no full text of the sourcecites [12] source
• Partly supported

  Section 2.3 describes partitioning a neural network between an edge device and the cloud, not dispatching a task across separate end, edge, and cloud tiers. Section 3.1 reports little accuracy loss and faster inference for some tested networks, but its table shows speed-ups below 1× for others. The source therefore does not support a general claim that this arrangement reduces execution delay. (LLM, conf=0.88)

  cites [49] sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [50] source

In the paper

“Besides, in order to optimize performance and efficiency, [108] presents an edge service framework, i.e., EdgeEye, which realizes a high-level abstraction of real-time video analytic functions based on DL.”

What the sources say

• Supported

  The Abstract identifies EdgeEye as an edge service framework for real-time video analytics and says it provides a high-level abstraction of video analysis functions based on DNNs. It also states that EdgeEye uses optimized inference engines for performance and efficiency. (LLM, conf=0.95)

  cites [108] source

In the paper

“In this way, the multi-capacity model provides variable resourceaccuracy trade-offs with a compact memory footprint, hence ensuring efficient multi-tenant DL at the end level.”

What the sources say

• Supported

  The Introduction states that NestDNN’s nested models provide varied resource–accuracy trade-offs with a compact memory footprint. It also describes NestDNN as enabling efficient multi-tenant on-device deep learning; the source supports the claim’s reference to the end-device level. (LLM, conf=0.93)

  cites [105] source

In the paper

“The main idea of this system is partitioning the DL model, used for inspection, and deploying them on the end, edge and cloud layer separately for improving the inspection efficiency.”

What the sources say

• Partly supported

  The Deep Model section (p. 4) says cameras capture images, fog nodes perform lower-level inference, and cloud servers complete inference when needed. The Offloading Strategy section (p. 6) describes allocating model layers to fog devices and cloud servers to improve efficiency. It does not describe deploying a separate part of the model on an end-device layer, so the claimed three-layer deployment is not supported. (LLM, conf=0.88)

  cites [113] sourceExperimental — accuracy not yet measured

In the paper

“In order to follow these principles, intelligent manufacturing should first address response latency, risk control, and privacy protection, and hence requires DL and edge computing.”

What the sources say

• Partly supported

  The Introduction (p. 2) describes response latency as a concern and proposes deep learning with fog computing for real-time manufacturing inspection. It does not establish that intelligent manufacturing generally requires these technologies, or identify risk control and privacy protection as reasons for adopting them. (LLM, conf=0.88)

  cites [113] sourceExperimental — accuracy not yet measured

In the paper

“For autonomous manufacturing inspection, DeepIns [113] uses DL and edge computing to guarantee performance and process delay respectively.”

What the sources say

• Supported

  Section III (p. 4) describes DeepIns as a deep model designed for strong classification performance in a fog-computing environment. The Introduction (p. 2) says its fog-side early exit reduces response latency, and the conclusion (p. 8) describes the system as real-time and efficient. (LLM, conf=0.90)

  cites [113] source

In the paper

“this solution can decompose the whole program at runtime instead of manually partitioning of programmers before program deploying.”

What the sources say

• ?Could not verify

  The Introduction (p. 2) and Federated Dropout section (p. 4) discuss reducing federated-learning model size and client computation, not decomposing a program at runtime. Searches for MAUI, runtime program partitioning, and pre-deployment manual partitioning found no passage addressing the claim. (LLM, conf=0.76)

  cites [178] sourceExperimental — accuracy not yet measured

In the paper

“The RAN controllers deployed at edge nodes are introduced in [180] to collect the data and run DL services, while the network controller, placed in the cloud, orchestrates the operations of the RAN controllers.In this manner, after running and feeding analytics and extract relevant metrics to DL models, these controllers can provide DL services to the users at the network edge.”

What the sources say

• Partly supported

  The Contributions section (p. 2) describes edge-deployed RAN controllers collecting data and running an ML infrastructure, with a cloud controller orchestrating them and the architecture providing new services to users. The source says machine learning, however, not specifically deep learning (DL); its listed techniques are regressors. The claim’s DL-specific description is therefore unsupported. (LLM, conf=0.88)

  cites [180] Polese, 2018 sourceExperimental — accuracy not yet measured

In the paper

“Communication expansion: 1.73×-1.98×”

What the sources say

• ?Could not verify

  Sections 3.3–3.4 (pp. 4–5) report computational cost and parameter counts, not communication expansion. Searches for communication overhead and the 1.73×–1.98× range found no passage addressing the claim. (LLM, conf=0.76)

  cites [222] sourceExperimental — accuracy not yet measured

In the paper

“q-FFL minimizes an aggregate re-weighted loss parameterized by q, so that devices with higher loss are given higher relative weight, thus encouraging less variance (i.e., more fairness) in the accuracy distribution.”

What the sources say

• Supported

  The Introduction (p. 1) states that q-FFL minimizes a q-parameterized, reweighted aggregate loss that gives higher-loss devices greater relative weight. Section 4.2 (p. 7) reports that q > 0 produces fairer testing-accuracy distributions with lower variance. (LLM, conf=0.95)

  cites [221] Li, 2019 source
• ?Not checked — no full text of the sourcecites [226] source

In the paper

“Downlink reduction: 14×; Uplink reduction: 28×; Local computation reduction: 1.7×”

What the sources say

• ?Could not verify

  The Introduction (p. 1) discusses reducing storage and computational costs, and Quantization and Binarization (p. 2) describes model compression. Neither passage reports downlink, uplink, and local-computation reductions of 14×, 28×, and 1.7×. Searches for those terms and figures found no passage addressing the claimed result. (LLM, conf=0.73)

  cites [205] Caldas, 2018 sourceExperimental — accuracy not yet measured

In the paper

“Experimental results corroborate Federate-dAveraging is robust to unbalanced and non-IID data and can facilitate the training process, viz., reducing the rounds of communication needed to train a DL model.”

What the sources say

• Supported

  The Introduction (p. 1) explicitly says FederatedAveraging is robust to unbalanced and non-IID data and reduces the communication rounds needed to train a deep network on decentralized data. The Experimental Results (p. 6) report speedups on non-IID and unbalanced data, supporting the citing sentence’s reference to experimental results. (LLM, conf=0.94)

  cites [198] McMahan, Communication-efficient learning of deep networks from decentralized data source

In the paper

“Scalability improvement: up to 1.5e6 clients”

What the sources say

• Partly supported

  The Operational Profile (p. 9) supports the general scalability claim: the system handled approximately 10 million daily active devices. It does not report 1.5 million clients as a scalability limit or measured result; it separately reports up to 10,000 devices participating simultaneously. (LLM, conf=0.85)

  cites [199] Bonawitz, 2019 sourceExperimental — accuracy not yet measured

In the paper

“Top 1 accuracy imrpovement: 5.59%-5.89%;Communication traffic reduction: 92%”

What the sources say

• Partly supported

  The Conclusion (p. 8) reports top-1 accuracy gains of 5.59% on imbalanced EMNIST and 5.89% on imbalanced CINIC-10 versus FedAvg. The Overhead section (p. 8), however, reports an 81.7% reduction in communication cost for its measured comparison, not 92%. (LLM, conf=0.91)

  cites [212] Duan, 2019 sourceExperimental — accuracy not yet measured

In the paper

“Adaptively minimizing q-FFL avoids the burden of handcrafting fairness constraints, and can adjust the goal according to the required fairness dynamically, achieving the effect of reducing the variance of accuracy distribution among partici-pated devices.”

What the sources say

• Supported

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

  cites [221] Li, 2019 source

In the paper

“First, DGC stipulates that only important gradients are exchanged, i.e., only gradients larger than a heuristically given threshold are transmitted.In order to avoid the information losing, the rest of the gradients are accumulated locally until they exceed the threshold.To be noted, gradients whether being immediately transmitted or accumulated for later exchanging will be coded and compressed, a”

What the sources say

• Supported

  Section 3.1 (p. 3) says DGC uses gradient magnitude as a heuristic, transmits gradients above a threshold, and accumulates the rest locally until they become large enough to send. It also describes encoding the sparse updates; Section 3.2 (p. 4) confirms that thresholded accumulated gradients are encoded and sent over the network. (LLM, conf=0.93)

  cites [191] Lin, 2017 source

In the paper

“Most of the gradient exchanges are redundant, and hence updated gradients can be compressed to cut down the communication cost while preserving the training accuracy (such as DGC in [191]).”

What the sources say

• Supported

  The Abstract (p. 1) says that 99.9% of gradient exchange in distributed SGD is redundant and reports that DGC reduces communication bandwidth without losing accuracy. The Introduction (p. 2) likewise reports compression of up to 600× without loss of accuracy. (LLM, conf=0.94)

  cites [191] Lin, 2017 source

In the paper

“Without requiring uploading data for central cloud training, FL [198], [199] can allow edge devices to train their local DL models with their own collected data and upload only the updated model instead.”

What the sources say

• Supported

  The Introduction (p. 1) says each client keeps its training dataset locally, computes an update to the server's global model, and communicates only that update. It also identifies local SGD on each client as the method for training deep networks without centrally storing the data. (LLM, conf=0.94)

  cites [198] McMahan, Communication-efficient learning of deep networks from decentralized data source
• Supported

  The Introduction (p. 1) says the system trains a deep neural network on data stored on a phone that never leaves the device, then combines the weights in the cloud. The operational profile (p. 13) specifies that devices upload updates to the global model, matching the claim’s distinction between uploading model updates and uploading training data. (LLM, conf=0.94)

  cites [199] Bonawitz, 2019 source

In the paper

“Communication rounds reduction: 10-100×”

What the sources say

• Supported

  In Experimental Results (p. 8), Table 3 reports FedAvg reaching accuracy targets in about 49–64 times fewer communication rounds than standard SGD. The same section reports 2,000 FedAvg rounds versus 197,500 SGD updates for similar test accuracy—approximately 99 times fewer. These experimental results support the stated 10–100× reduction. (LLM, conf=0.90)

  cites [198] McMahan, Communication-efficient learning of deep networks from decentralized data source

In the paper

“By using the cross-modal transfer in the learning of edge devices across different sensing modalities, required labeled data and the training process can be largely reduced and accelerated, respectively.”

What the sources say

• Supported

  The Abstract states that RecycleML uses cross-modal transfer across edge-device sensing modalities to reduce labeled-data needs and speed training. The Introduction reports these gains on the CMActivity activity-recognition dataset, and the Conclusion likewise reports reduced training time. (LLM, conf=0.94)

  cites [284] source

In the paper

“Existing FL methods [198], [199] focus on synchronous training, and can only process hundreds of devices in parallel.”

What the sources say

• Partly supported

  The Introduction (p. 2) says this paper assumes a synchronous update scheme, supporting that characterization of its method. It does not say existing federated-learning methods can only process hundreds of devices in parallel; the 100-client figure in the Experimental Results (p. 5) describes one experimental setup, not a capacity limit. (LLM, conf=0.87)

  cites [198] McMahan, Communication-efficient learning of deep networks from decentralized data sourceExperimental — accuracy not yet measured
• Partly supported

  The Introduction (p. 1) says the authors chose to support synchronous rounds, and Future Work (p. 10) says current FL algorithms can efficiently utilize only hundreds of devices in parallel. The claim drops the efficiency qualifier: the Operational Profile (p. 9) reports up to 10,000 devices participating simultaneously, so the paper does not establish a hard limit of hundreds that can be processed. (LLM, conf=0.87)

  cites [199] Bonawitz, 2019 sourceExperimental — accuracy not yet measured

In the paper

“The application of vanilla FL is using RNN for next-word-prediction [199], in which the training data can be obtained along with users' daily inputs.”

What the sources say

• Supported

  The Applications section (p. 8) says Gboard used the FL platform to train an RNN for next-word prediction. It also says FL training data can be generated on devices, with labels inferred from user activity such as typed words, supporting the description of data obtained from users’ inputs. (LLM, conf=0.91)

  cites [199] Bonawitz, 2019 source

In the paper

“To solve this kind of task offloading problem is NP-hard [244], since at least combination optimization of communication and computing resources along with the contention of edge devices is required.”

What the sources say

• Partly supported

  The Introduction (p. 2) says finding a centralized optimal multi-user offloading solution is NP-hard in a multi-channel wireless interference environment, and that the model accounts for communication and computation. But it treats wireless contention as a separate extension; the cited passages do not establish that contention is required for, or explains, the NP-hardness claim. Related Work (p. 12) also limits the result to the generalized multi-channel problem, noting that the single-channel case is not NP-hard. (LLM, conf=0.85)

  cites [244] sourceExperimental — accuracy not yet measured

In the paper

“Privacy: the data required for DL might carry a lot of private information, and privacy issues are critical to areas such as smart home and cities.”

In the paper

“a large number of computation tasks need to be delivered to the cloud for processing [3], which undoubtedly poses serious challenges on network capacity and the computing power of cloud computing infrastructures;”

What the sources say

• ?Not checked — no full text of the sourcecites [3] source

In the paper

“3) the hierarchical computing architecture provides more reliable DL computation;”

In the paper

“Reliability: most cloud computing applications relies on wireless communications and backbone networks for connecting users to services, but for many industrial scenarios, intelligent services must be highly reliable, even when network connections are lost;”

In the paper

“Cost: training and inference of DL models in the cloud requires devices or users to transmit massive amounts of data to the cloud, thus consuming a large amount of network bandwidth;”

In the paper

“4) with richer data and application scenarios, edge computing can promote the pervasive application of DL and realize the prospect of”

What the sources say

• ?Not matched

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

  cites [13] 2019

In the paper

“Latency: the delay to access cloud services is generally not guaranteed and might not be short enough to satisfy the requirements of many time-critical applications such as cooperative autonomous driving [11];”

What the sources say

• ?Not checked — no full text of the sourcecites [11] source

In the paper

“In fact, edge computing is gradually being combined with Artificial Intelligence (AI), benefiting each other in terms of the realization of edge intelligence and intelligent edge as depicted in Fig. 1.”

In the paper

“Since the edge is closer to users than the cloud, edge computing is expected to solve many of these issues.”

In the paper

“on one hand, edge intelligence is expected to push DL computations from the cloud to the edge as much as possible, thus enabling various distributed, lowlatency and reliable intelligent services.”

In the paper

“3) powerful cloud backup, the cloud can provide powerful processing capabilities and massive storage when the edge cannot afford.”

In the paper

“2) agile service response, services hosted at the edge can significantly reduce the delay of data transmissions and improve the response speed;”

In the paper

“According to a Cisco white paper [1], 50 billion IoT devices will be connected to the Internet by 2020.”

What the sources say

• ?Not matched

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

  cites [1] 2019

In the paper

“5) diversified and valuable DL services can broaden the commercial value of edge computing and accelerate its deployment and growth.”

In the paper

“Cisco estimates that nearly 850 Zettabytes (ZB) of data will be generated each year outside the cloud by 2021, while global data center traffic is only 20.6 ZB [2].”

What the sources say

• ?Not matched

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

  cites [2] 2019

In the paper

“With respect to CV, NLP, and AI, DL is adopted in a myriad of applications and corroborates its superior performance [78].”

What the sources say

• ?Not checked — no full text of the sourcecites [78] source

In the paper

“Unfortunately, executing these tasks in the cloud often incurs high bandwidth consumption, unexpected latency, and reliability issues.”

In the paper

“Edge nodes are expected to have computing and caching capabilities and to provide high-quality network connection and computing services near end devices.”

In the paper

“Nonetheless, the edge computing architecture, on account of it covers a large number of distributed edge devices, can be utilized to better serve DL.”

In the paper

“Currently, a large number of GPUs, TPUs, or FPGAs are required to be deployed in the cloud to process DL service requests.”

In the paper

“For DL services with complex configuration and intensive resource requirements, edge computing systems with advanced and excellent microservice architecture are the future development direction.”

In the paper

“Software-defined Networking (SDN) allows rapid deployment of services, network programmability and multi-tenancy support, through three key innovations [72]: 1) Decoupling of control planes and data planes; 2) Centralized and programmable control planes; 3) Standardized application programming interface.”

What the sources say

• ?Not checked — no full text of the sourcecites [72] source

In the paper

“For example, LAVEA [106] attaches edge nodes to the same access point or BS as well as the end devices, which ensure that services can be as ubiquitous as Internet access.”

What the sources say

• ?Not checked — no full text of the sourcecites [106] source

In the paper

“There are two common ways to perform distributed training, i.e., data parallelism and model parallelism [95]- [98] as illustrated in Fig. 10.”

What the sources say

• ?Not matched

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

  cites [95]- [98]

In the paper

“Typical network optimization methodologies rely on fixed mathematical models; however, it is difficult to accurately model rapidly changing edge network environments and systems.”

In the paper

“In general, VM is better at isolating while container provides easier deployment of repetitive tasks [69].”

What the sources say

• ?Not checked — no full text of the sourcecites [69] source

In the paper

“The focus of this collaborative paradigm is not only the successful completion of tasks but also achieving the optimal balance of equipment energy consumption, server loads, transmission and execution delays.”

In the paper

“Coincidentally, a large number of end devices, edge nodes, and cloud data centers, are scattered and envisioned to be connected by virtue of edge computing networks.These distributed devices can potentially be powerful contributors once the DL training jumps out of the cloud.”

In the paper

“The edge computing architecture will become an important complement to the cloud, even replacing the role of the cloud in some scenarios.”

In the paper

“Edge computing cannot provide that resources for DL services as the cloud does.”

In the paper

“To fully exploit the bond function of the edge, VideoEdge [109] implements an end-edge-cloud hierarchical architecture to help achieve load balancing concerning analytical tasks while maintaining high analysis accuracy.”

What the sources say

• ?Not checked — no full text of the sourcecites [109] source

In the paper

“At the end level, video capture devices, such as smartphones and surveillance cameras are responsible for video capture, media data compression [102], image preprocessing, and image segmentation [103].”

What the sources say

• ?Not checked — no full text of the sourcecites [102] source
• ?Not checked — no full text of the sourcecites [103] source

In the paper

“Training a large DL model generally requires a lot of computation resources, even thousands of CPUs are required to train a large-scale DL model.”

In the paper

“The combination of edge computing and DL to form high-performance Edge DL services requires the coordinated integration of computing, networking and communication resources, as depicted in Fig. 8.”

In the paper

“Certainly, edge devices typically have limited computing power or power consumption compared to the cloud.”

In the paper

“Container virtualization for edge computing can effectively reduce the workload execution time with high performance and storage requirements, and can also deploy a large number of services in a scalable and straightforward fashion [70].”

What the sources say

• ?Not checked — no full text of the sourcecites [70] source

In the paper

“In addition, edge nodes also have the ability to cache, which can improve the response time by caching popular contents.”

In the paper

“A container consists of a single file that includes an application and execution environment with all dependencies, which makes it enable efficient service handoff to cope with user mobility [71].”

What the sources say

• ?Not checked — no full text of the sourcecites [71] source

In the paper

“Therefore, by deploying edge nodes to perform the computation task, the task processing can be accelerated while ensuring accuracy.”

In the paper

“Further, Edge DL services typically require high bandwidth, low latency, and dynamic network configuration, while Software-defined Networking (SDN) allows rapid deployment of services, network programmability and multi-tenancy support, through three key innovations [72]:”

In the paper

“For example, with the support of NFV/SDN, edge nodes can be efficiently orchestrated and integrated with cloud data centers [73].”

What the sources say

• ?Not checked — no full text of the sourcecites [73] source

In the paper

“However, the edge computing community has not yet reached a consensus on the standardized definitions, architectures and protocols of edge computing [23].”

What the sources say

• ?Not checked — no full text of the sourcecites [23] source

In the paper

“In general, applying DL for it requires high computation and storage resources.”

In the paper

“Since the distributed model training performance on an edge node may be significantly impaired due to its local knowledge, the cloud needs to integrate different well-trained DL models to achieve global knowledge.”

What the sources say

• ?Not checked — no full text of the sourcecites [102] source

In the paper

“In this section, as depicted in Fig. 11, we summarize related works as a hybrid hierarchical architecture, which is divided into three levels: end, edge, and cloud.”

In the paper

“DL is expected to deal with this problem: when faced with complex and cumbersome network information, DL can rely on its powerful learning and reasoning ability to extract valuable information from data and make adaptive decisions, achieving intelligent maintenance and management accordingly.”

In the paper

“Therefore, the combination of DL and edge computing is not straightforward and requires a comprehensive understanding of DL models and edge computing features for design and deployment.”

In the paper

“At the cloud level, the cloud is responsible for the integration of DL models among the edge layer and updating parameters of distributed DL models on edge nodes [102].”

What the sources say

• ?Not checked — no full text of the sourcecites [102] source

In the paper

“Nonetheless, distributed training can facilitate the training process by taking full advantage of parallel servers.”

In the paper

“When the edge is unable to provide the service confidently (e.g., detecting objects with low confidence), the cloud can use its powerful computing power and global knowledge for further processing and assist the edge nodes to update DL models.”

In the paper

“If these operations can largely reduce the size of intermediate data and hence reduce the network traffic, the pressure of backbone networks can be alleviated.”

In the paper

“customized edge DL frameworks can help efficiently 1) match edge platform and DL models; 2) exploit underlying hardware in terms of performance and power; 3) orchestrate and maintain DL services automatically.”

What the sources say

• ?Not checked — no full text of the sourcecites [179] source

In the paper

“For the protection of sensitive home data, the data processing of smart home systems must rely on edge computing.”

In the paper

“Have a speedup ranging from 1.3× to 4× (1.2× to 1.7×) against running in local (clientcloud confguration)”

What the sources say

• ?Not checked — no full text of the sourcecites [106] source

In the paper

“improve throughput by 1.7×-2.2×and speedup CNN inference by 1.7×-3.5×”

What the sources say

• ?Not checked — no full text of the sourcecites [156] source

In the paper

“Reduce memory footprint by more than 68% without sacrificing accuracy”

What the sources say

• ?Not checked — no full text of the sourcecites [156] source

In the paper

“Besides, Vehicle-to-Vehicle (V2V) communication technology can be taken advantaged to further connect vehicles, either as an edge node or an end device managed by DRL-based control policies [112].”

What the sources say

• ?Not checked — no full text of the sourcecites [112] source

In the paper

“Since this system involves multi-dimensional control, a DRLbased approach is first utilized to solve the optimization problem for enhancing the holistic system performance.”

What the sources say

• ?Not checked — no full text of the sourcecites [110] source

In the paper

“IONN divides a DL model, prepared for uploading, into multiple partitions, and uploads them to the edge node in sequential.The edge node, receiving the partitioned models, incrementally builds the DL model as each partitioned model arrives, while being able to execute the offloaded partial DL computation even before the entire DL model is uploaded.”

What the sources say

• ?Not checked — no full text of the sourcecites [175] source

In the paper

“A natural choice of collaboration is the edge performs data pre-processing and preliminary learning, when the DL tasks are offloaded.Then, the intermediate data, viz., the output of edge architectures, are transmitted to the cloud for further DL computation [176].”

What the sources say

• ?Not checked — no full text of the sourcecites [176] source

In the paper

“Besides, the power, energy efficiency, memory footprint limitation of IoT edge devices [116] should also be considered.”

What the sources say

• ?Not checked — no full text of the sourcecites [116] source

In the paper

“Like use cases in [118], [119], edge computing is deployed to optimize indoor positioning systems and home intrusion monitoring so that they can get lower latency than using cloud computing as well as the better accuracy.”

What the sources say

• ?Not checked — no full text of the sourcecites [118] source
• ?Not checked — no full text of the sourcecites [119] source

In the paper

“In [12], all layers of a DNN are profiled on the end device and the edge node in terms of the data and computation characteristics, in order to generate performance prediction models.Based on these prediction models, wireless conditions and server load levels, the proposed Neurosurgeon evaluates each candidate point in terms of end-to-end latency or mobile energy consumption and partition the DNN”

What the sources say

• ?Not checked — no full text of the sourcecites [12] source

In the paper

“Further, the combination of DL and edge computing can make these intelligent services become more various and powerful.”

In the paper

“DL tasks are usually computationally intensive and requires large memory footprints.”

In the paper

“In intelligent factories, edge computing is conducive to expand the computation resources, the network bandwidth, and the storage capacity of the cloud to the IoT edge, as well as realizing the resource scheduling and data processing during manufacturing and production [114].”

What the sources say

• ?Not checked — no full text of the sourcecites [114] source

In the paper

“In contrast to only partitioning the DNN by layers as in [12], FTP can fuse layers and partitions them vertically in a grid fashion, hence minimizing the required memory footprint of participated edge devices regardless of the number of partitions and devices, while reducing communication and task migration cost as well.”

What the sources say

• ?Not checked — no full text of the sourcecites [156] source

In the paper

“The original intention of applying edge computing in smart cities is more due to cost and efficiency considerations.”

In the paper

“ALOHA [182] formulates a toolflow: 1) Automate the model design.It generates the optimal model configuration by taking into account the target task, the set of constraints and the target architecture; 2) Optimize the model configuration.It partitions the DL model and accordingly generates architecture-aware mapping information between different inference tasks and the available resources.3) Autom”

What the sources say

• ?Not checked — no full text of the sourcecites [182] source

In the paper

“Nevertheless, the intermediate output, e.g., high-resolution surveillance video streams, should be carefully designed much smaller than the raw input, therefore drastically reducing the network traffic required between the end and the edge (or the edge and the cloud).”

In the paper

“Therefore, DL models should be further customized to fit in the resource-constrained edge, while carefully treating the trade-off between the inference accuracy and the execution latency of them.”

In the paper

“Certainly, the outstanding performance of DL models is inseparable from the support of high-level hardware, and it is difficult to deploy them in the edge with limited resources.”

In the paper

“Nonetheless, with the exponential growth of IoT edge devices, 1) how to remotely manage evolving DL models and 2) how to continuously evaluate these models for them are necessary.”

In the paper

“It is envisioned that vehicles can be connected to improve safety, enhance efficiency, reduce accidents, and decrease traffic congestion in transportation systems [110].”

What the sources say

• ?Not checked — no full text of the sourcecites [110] source

In the paper

“MoDNN, proposed in [177], executes DL in a local distributed mobile computing system over a Wireless Local Area Network (WLAN).Each layer of DNNs is partitioned into slices to increase parallelism and to reduce memory footprint, and these slices are executed layerby-layer.By the execution parallelism among multiple end devices, the DL computation can be significantly accelerated.”

What the sources say

• ?Not checked — no full text of the sourcecites [177] source

In the paper

“Moreover, for important data sources, data safety and privacy protection should be addressed.”

In the paper

“For instance, it endows robots the ability of dynamic visual servicing [120] and enables efficient music cognition system [121].”

What the sources say

• ?Not checked — no full text of the sourcecites [120] Tian, 2018 source
• ?Not checked — no full text of the sourcecites [121] source

In the paper

“The natural characteristic of geographically distributed data sources in cities requires an edge computingbased paradigm to offer location-awareness and latencysensitive monitoring and intelligent control.”

In the paper

“On one hand, edge computing provides low-latency, highspeed communication and fast-response services for vehicles, making automatic driving possible.”

In the paper

“This architecture can not only support latency-sensitive applications on end devices but also perform slightly latency-tolerant tasks efficiently on edge nodes, while large-scale DL models responsible for deep analysis are hosted on the cloud.”

What the sources say

• ?Not checked — no full text of the sourcecites [122] source

In the paper

“FedPAQ [204] simultaneously incorporates these features and provides near-optimal theoretical guarantees for both strongly convex and non-convex loss functions, while empirically demonstrating the communication-computation tradeoff.”

What the sources say

• ?Not checked — no full text of the sourcecites [204] Reisizadeh, 2019 source

In the paper

“To summarize, FL can deal with several key challenges in edge computing networks: 1) Non-IID training data.”

In the paper

“Zoo [184] provides a concise Domain-specific Language (DSL) to enable easy and type-safe composition of DL services.”

What the sources say

• ?Not checked — no full text of the sourcecites [184] source

In the paper

“ECO [185] uses a graph-based overlay network approach to 1) model and track pipelines and dependencies and then 2) map them to geographically distributed analytic engines ranging from small edge-based engines to powerful multi-node cloud-based engines.By this means, DL computation can be distributed as needed to manage cost and performance, while also supporting other practical situations, such as”

What the sources say

• ?Not checked — no full text of the sourcecites [185] Talagala, ECO: Harmonizing edge and cloud with ML/DL orchestration

In the paper

“Training acceleration: 2×; Model accuracy improvement: 4%”

What the sources say

• ?Not checked — no full text of the sourcecites [218] Xu, 2019 source

In the paper

“Besides, in practical, edge devices may suffer from higher latency, lower transmission rate and intermittent connections, and therefore further hindering the gradients exchanging between DL models belong to different edge devices.”

In the paper

“Naturally, the edge architecture, which consists of a large number of edge nodes with modest computing resources, can cater for alleviating the pressure of networks by processing the data or training at themselves.”

In the paper

“Experimental results show that no single DL library or hardware platform can entirely outperform others, and loading the DL model may take more time than that of executing it.”

What the sources say

• ?Not checked — no full text of the sourcecites [186] Zhang, pCAMP: Performance comparison of machine learning packages on the edges source
• ?Not checked — no full text of the sourcecites [54] Ignatov, 2018 source

In the paper

“Communication cost reduction: by two orders of magnitude”

What the sources say

• ?Not checked — no full text of the sourcecites [203] Konečný, 2016 source

In the paper

“In addition, merging all data into the cloud might violate privacy issues.”

In the paper

“Nonetheless, FL is emerging and is promised to address these issues.”

In the paper

“However, this proposed method is designed for seismic imaging application and can not be generalized for future DL training, since the communication cost for training large scale DL models is extremely high.”

What the sources say

• ?Not checked — no full text of the sourcecites [189] source

In the paper

“Besides, a significant amount of communication resources will be consumed, and hence aggravating wireless and backbone networks if massive data are required to be continually transmitted from distributed end devices or edge nodes to the cloud.”

In the paper

“FL must be robust to the unexpectedly drop out of edge devices, or else once a device loses its connection, the synchronization of FL in one round will be failed.”

In the paper

“Leverage blockchain to exchange and verify model updates of local training Learning completion latency”

What the sources say

• ?Not checked — no full text of the sourcecites [223] Kim, 2018 source

In the paper

“Therefore, in [223], combining Blockchain and FL as BlockFL is proposed to realize 1) locally global model updating at each edge device rather a specific server, ensuring device malfunction cannot affect other local updates when updating the global model; 2) appropriate reward mechanism for stimulating edge devices to participate in FL.”

What the sources say

• ?Not checked — no full text of the sourcecites [223] Kim, 2018 source

In the paper

“Hence, Laplace noises [197] are added to these possibly exposed training data for enhancing the training data privacy assurance.”

What the sources say

• ?Not checked — no full text of the sourcecites [196] source
• ?Not checked — no full text of the sourcecites [197] source

In the paper

“However, the server also should not trust edge devices completely, since devices with abnormal behavior can forge or poison their training data, which results in worthless model updates, and hence harming the global model.”

In the paper

“Convergence acceleration: 62.4%”

What the sources say

• ?Not checked — no full text of the sourcecites [211] Hu, 2019 source

In the paper

“Specifically, the total wall-clock time of FL includes not only the computation time but also the communication time of all clients.”

In the paper

“OpenEI [183] defines each DL algorithm as a four-element tuple ¡Accuracy, Latency, Energy, Memory Footprint¿ to evaluate the Edge DL capability of the target hardware platform.Based on such tuple, OpenEI can select a matched model for a specific edge platform based on different Edge DL capabilities in an online manner.”

What the sources say

• ?Not checked — no full text of the sourcecites [183] Zhang, 2019 source

In the paper

“Nonetheless, a standard testbed for Edge DL is missing, which hinders the study of edge architectures for DL.”

In the paper

“Though largescale distributed training at edge evades transmitting bulky raw dataset to the cloud, the communication cost for gradients exchanging between edge devices is inevitably introduced.”

In the paper

“Use the trimmed mean as a robust aggregation Top 1 accuracy against data poisoning”

What the sources say

• ?Not checked — no full text of the sourcecites [201] Xie, 2019 source

In the paper

“Compared to DQL, the experiment results indicate that Double-DQL can save more energy and achieve higher training efficiency.”

What the sources say

• ?Not checked — no full text of the sourcecites [257] source

In the paper

“Though offloading computation tasks to edge nodes can enhance the processing efficiency of the computation tasks, the reliability of offloading suffers from the potentially low quality of wireless environments.”

In the paper

“In this regard, there are many potential areas where DL can be applied, including online content streaming [285], routing and traffic control [286] [287], etc.”

What the sources say

• ?Not checked — no full text of the sourcecites [285] source
• ?Not checked — no full text of the sourcecites [286] source
• ?Not checked — no full text of the sourcecites [287] source

In the paper

“In summary, the key issue of extending DL from the cloud to the edge of the network is: under the multiple constraints of networking, communication, computing power, and energy consumption, how to devise and develop edge computing architecture to achieve the best performance of DL training and inference.”

In the paper

“Therefore, the performance of DNN-based methods bounds by fixed optimization algorithms and is not self-adapted.”

In the paper

“DRL-based security solutions can provide secure offloading (from the edge device to the edge node) to against jamming attacks [261] or protect user location privacy and the usage pattern privacy [270].”

What the sources say

• ?Not checked — no full text of the sourcecites [261] source
• ?Not checked — no full text of the sourcecites [270] source

In the paper

“However, since DL solutions do not rely entirely on accurate modeling of networks and devices, finding a scenario where DL can be applied is not the most important concern.”

In the paper

“Among all these works related to learning-based optimization methods, DL-based approaches have advantages over others when multiple edge nodes and radio channels are available for computation offloading.”

In the paper

“Hence, taking the limited computing capabilities of edge devices (at least for now) into consideration, partitioning a large-scale DL model and allocating these segments to different edge devices for training may be a more feasible and practical solution.”

In the paper

“Besides, if applying DL to optimize realtime edge computing networks, the training and inference of DL models or DRL algorithms may bring certain side effects, such as the additional bandwidth consumed by training data transmission and the latency of DL inference.”

In the paper

“If the number of surviving devices is too small, the server has to discard the entire epoch including all received updates.”

In the paper

“Some devices may no longer be available when they should be synchronized, and hence the server must determine the timeout threshold to discard the laggard.”

In the paper

“Since edge devices generally equipped with limited computation, energy and radio resources, the transmission between them and the edge node is more vulnerable to various attacks, such as jamming attacks and Distributed Denial of Service (DDoS) attacks, compared to cloud computing.”

In the paper

“Therefore, edge management and maintenance require optimizations from multiple perspectives (including communication perspective).”

In the paper

“Edge computing allows edge devices offload part of their computing tasks to the edge node [243], under constraints of energy, delay, computing capability, etc.”

What the sources say

• ?Not checked — no full text of the sourcecites [243] source

In the paper

“To perform more fine-grained power control for local execution and task offloading, policy-gradient-based DRL should be considered.”

In the paper

“Traditional RL algorithms are limited by the requirement for handcrafting features and the flaw that hardly handling high-dimensional observation data and actions [239].”

What the sources say

• ?Not checked — no full text of the sourcecites [239] source

In the paper

“For example, end devices, edge nodes, and the cloud can be deemed as individual agents.By this means, each agent trains its own strategy according to its local imperfect observations, and all participated agents work together for optimizing edge computing networks.”

In the paper

“However, differing from parallel training over bus-or switch-connected CPUs or GPUs in the cloud, perform model training at distributed edge devices should further consider wireless environments, device configurations, privacies, etc.”

In the paper

“DL, as a key technique of artificial intelligence, and edge computing are expected to benefit each other.”