Convergence of Edge Computing and Deep Learning: A Comprehensive Survey

Motivation

• Cloud-centric DL is strained by distributed data, high transmission costs, uncertain latency, reliability requirements, and privacy concerns.These constraints are especially relevant to time-critical, industrial, smart-home, and smart-city applications.Source: §I · pp. 1–2
• Edge intelligence moves DL computation toward users and data sources to support distributed, low-latency, reliable services while reducing cloud communication and improving privacy.Local storage and hierarchical computation can limit raw-data transmission to the cloud.Source: §I · p. 2
• Intelligent edge incorporates DL into edge infrastructure for adaptive maintenance and management of complex, rapidly changing communication and computing environments.DL extracts information from network data and supports adaptive decisions where fixed mathematical models are difficult to apply accurately.Source: §I · p. 2

Scope and taxonomy

• The survey identifies five Edge DL technologies: applications on edge, inference in edge, edge computing for DL, training at edge, and DL for optimizing edge.Together they cover service deployment, model execution, platform support, distributed training, and functions such as caching and computation offloading.Source: §I · pp. 2–3
• Its taxonomy links DL applications on edge and DL for optimizing edge to edge intelligence and intelligent edge, respectively, with training and inference providing supporting capabilities.The survey emphasizes that deploying DL on the edge alone is incomplete because DL and edge computing benefit each other.Source: §I · p. 3

Survey organization

• The paper surveys fundamentals, five enabling technologies, lessons learned, and open challenges concerning the convergence of DL and edge computing.The organization spans DL applications, inference, edge platforms, training, and optimization.Source: §I · p. 3

• Edge computing reduces data transmission, improves service latency, and eases pressure on cloud computing.Source: §II · p. 4

Cloudlet and Micro Data Centers

• Cloudlets combine mobile and cloud computing as a middle layer between mobile devices and the cloud.Source: 1) Cloudlet and Micro Data Centers · p. 4
• Micro data centers package computing, storage, and networking equipment in one enclosure for applications needing lower latency or serving resource-limited devices.Source: 1) Cloudlet and Micro Data Centers · p. 4

• Fog computing uses a distributed multi-tier architecture targeted to geographic areas and real-time applications requiring reduced latency.Source: 2) Fog Computing · p. 4
• Multi-access Edge Computing places computing capabilities and service environments at cellular-network edges, providing lower latency, context awareness, and higher bandwidth.Source: 3) Mobile (Multi-access) Edge… · p. 4

Definition of Edge Computing Terminologies

• The paper distinguishes end devices, such as smartphones and smart vehicles, from edge nodes, such as Cloudlets, RSUs, fog nodes, and edge servers.Source: 4) Definition of Edge Computing… · p. 4

AI Hardware for Edge Computing

• GPU-, FPGA-, and ASIC-based hardware offer different trade-offs in compatibility, performance, energy use, computation resources, and programming flexibility.Source: 1) AI Hardware for Edge Computing · p. 5
• Smartphone AI chips increasingly support DL acceleration, including Qualcomm Snapdragon hardware with an SDK supporting major DL frameworks.Source: 1) AI Hardware for Edge Computing · p. 5
• Edge nodes provide greater computing capability than most end devices and can respond faster than the cloud, accelerating task processing while maintaining accuracy.Source: 2) Integrated Commodities Potentially… · p. 6

Virtualizing the Edge

• Virtualization must maximize limited edge resources and isolate complex DL software libraries with their versions and dependencies.Source: C. Virtualizing the Edge · p. 6

Virtualization Techniques

• VMs provide stronger isolation, whereas containers simplify deployment, reduce execution overhead, and support scalable service handoff.Source: 1) Virtualization Techniques · p. 6

Network Virtualization

• NFV and SDN improve flexible edge-network management by separating network functions from dedicated hardware and enabling programmable, rapidly deployed services.Source: 2) Network Virtualization · pp. 6–7

Network Slicing

• Network slicing creates multiple service-optimized network instances over shared physical infrastructure.Source: 3) Network Slicing · p. 7
• Coordinating network slicing with computing and communication optimization supports distributed edge-computing paradigms and customized resource allocation.Source: 3) Network Slicing · p. 7
• Virtualization and SDN jointly support service customization, flexible control, and resource provision for Edge DL services.Source: 3) Network Slicing · p. 7

Deep Neural Networks

• Fully connected neural networks feed each layer forward but have high complexity, modest performance, and slow convergence.Source: 1) Fully Connected Neural Network (FCNN) · p. 8
• CNNs use pooling and moving filters to capture local correlations, extract features, reduce model complexity, and mitigate overfitting.Source: 3) Convolutional Neural Network (CNN) · p. 8

Recurrent Neural Network

• LSTMs add gates and memory cells to RNNs, controlling information flow to overcome gradient explosion.Source: 5) Recurrent Neural Network (RNN) · p. 8

Transfer Learning and Knowledge Distillation

• Transfer learning and knowledge distillation reuse knowledge from trained models to accelerate training or reduce inference overhead.Source: 6) Transfer Learning (TL) · p. 8

Deep Reinforcement Learning

• DRL uses DNN representation ability to fit value functions or policies for decision-making in large or continuous state-action spaces.Source: B. Deep Reinforcement Learning (DRL) · p. 8

Value-Based DRL

• Value-based DRL methods such as DQL use DNNs to map high-dimensional inputs to actions, with replay and target networks stabilizing training.Source: 1) Value-based DRL · p. 8

Policy-Gradient DRL

• Policy-gradient methods optimize policy parameters using gradients of expected reward, while Actor-Critic assigns policy and value estimation to separate networks.Source: 2) Policy-gradient-based DRL · pp. 8–9

DL Training at Edge

• Distributed training uses data or model parallelism to exploit parallel servers and address the time and resource costs of centralized training.Source: C. Distributed DL Training · p. 9
• End, edge, and cloud devices can contribute distributed resources when DL training moves beyond centralized cloud infrastructure.Source: C. Distributed DL Training · p. 9

DL Applications on Edge

• Edge deployment supports real-time DL services because cloud-only architectures may fail to meet latency requirements for analytics and smart manufacturing.Source: §IV · p. 9, A. Real-time Video Analytic · p. 9

Real-Time Video Analytics

• Video offloading must balance network conditions, data use, battery consumption, processing delay, frame rate, and analytics accuracy.Source: 1) End Level · pp. 9–10

Vehicular Networks

• Vehicular edge frameworks integrate networking, caching, and computation, using DRL to optimize holistic system performance and task offloading.The framework supports dynamic orchestration of multidimensional resources for different vehicular applications.Source: B. Autonomous Internet of Vehicles… · p. 10

Intelligent Manufacturing

• DeepIns partitions an inspection model across end, edge, and cloud layers to improve inspection efficiency while preserving process-delay performance.The system targets autonomous manufacturing inspection and distributes model components across the deployment hierarchy.Source: C. Intelligent Manufacturing · p. 10
• Intelligent manufacturing must manage evolving models and continuously evaluate them on IoT devices while coping with power, energy, and memory limitations.Caching, heterogeneous-device communication, and computation offloading are identified as ways to address these resource bottlenecks.Source: C. Intelligent Manufacturing · p. 10

Smart Homes and Cities

• Smart-home edge computing protects sensitive data while enabling applications such as dynamic visual servicing and music cognition.The paper also identifies geographically distributed, latency-sensitive monitoring and intelligent control as motivations for edge computing in smart cities.Source: D. Smart Home and City · pp. 10–11

Edge Model Constraints

• Large-scale DNNs are difficult to deploy at the edge because deeper and wider models require substantial computation, memory, and supporting hardware.Model optimization must reduce resource costs without significant accuracy loss, using approaches such as pruning, quantization, and low-rank factorization.Source: §V · p. 11, A. Optimization of DL Models in Edge · p. 11, 1) General Methods for Model… · p. 11

Optimization Overview

• Edge DL optimization must account for limited computing, memory, bandwidth, and power while preserving model accuracy.The section distinguishes optimization efforts according to their impact on inputs, structures, model choices, and execution frameworks.Source: 2) Model Optimization for Edge Devices · p. 11

Model Input

• Input optimization narrows application-specific search spaces by filtering non-target frames and tuning video resolution or sampling configurations online.FFS-VA uses specialized filters and a small detector, while Chameleon reduces the cost of searching input-stream configurations.Source: 2) Model Optimization for Edge Devices · p. 11, Figure 12 · p. 11

Model Structure

• Structure optimization uses specialized architectures, difference detectors, adaptive parameter pruning, and cross-layer algorithm–software–hardware optimization.These techniques target faster inference by exploiting model structure and application-specific redundancy.Source: 2) Model Optimization for Edge Devices · pp. 11–12, Figure 13 · p. 12

Model Selection

• Model selection weighs precision against inference time, using predictors and automatically tuned input features to choose among available or compressed models.Combining compression techniques can provide multiple model variants for adaptive selection at the edge.Source: 2) Model Optimization for Edge Devices · p. 12

Model Framework

• Model frameworks provide optimized kernels, compression, quantization, coding, and hardware-aware support for deploying computationally demanding DNNs on edge devices.The cited framework capabilities address both software execution and the reduction of model memory demands.Source: 2) Model Optimization for Edge Devices · p. 12

Distributed and Adaptive Inference

• Partitioning models across local processors, distributed edge nodes, or end–edge–cloud systems reduces cloud-upload bottlenecks, while early exits and caching further improve inference efficiency.EEoI sends only immediate features for uncertain cases and enables confident samples to finish locally; caching and stem sharing reduce redundant computation.Source: B. Segmentation of DL Models · p. 12, C. Early Exit of Inference (EEoI) · pp. 12–13

Edge Support for DL

• Extensive DL deployment requires customized edge hardware and software, computation offloading, service maintenance, and fair performance evaluation.These components collectively support efficient execution and continued development of Edge DL systems.Source: §VI · p. 13

Mobile CPUs and GPUs

• CMSIS-NN minimizes neural-network memory footprints on ARM Cortex-M cores, fitting DL models into IoT devices while maintaining normal performance and energy efficiency.The result addresses computation, memory, and energy constraints on lightweight edge devices.Source: 1) Mobile CPUs and GPUs · pp. 13–14
• Mobile DL training performance depends strongly on model size and efficient fusion of mobile CPUs and GPUs.The evaluation used a relatively small Mentee network because common models such as VGG exceed mainstream edge-device memory.Source: 1) Mobile CPUs and GPUs · p. 14

FPGA-based Solutions

• FPGA edge devices accelerate CNNs with reconfigurable operations and can outperform state-of-the-art CPU and GPU implementations in RNN speech recognition while using less energy.An FPGA-based offloading platform also showed energy-consumption and hardware-cost advantages over GPU- or CPU-based alternatives in vision applications.Source: 2) FPGA-based Solutions · p. 14
• Whether FPGAs or GPUs/CPUs are more suitable for edge computing remains unresolved because FPGA programming and model transplantation require additional work.FPGAs offer workload-insensitive throughput, high-concurrency performance, and energy efficiency, but their development ecosystem remains less familiar.Source: 2) FPGA-based Solutions · p. 14

Computation Offloading

• End-cloud, end-edge, and collaborative offloading modes trade implementation simplicity against task decomposition, scheduling, accuracy, latency, model size, battery, and network conditions.Tables IV and V organize solution comparisons and edge communication/computation modes, while edge nodes may be insufficient for large concurrent query volumes.Source: 1) Integral Offloading · pp. 14, 16, Table IV · p. 14, Table V · p. 15

Partial Offloading

• Partial offloading enables online fine-grained partitioning of DL tasks between end devices and edge nodes under network constraints.MAUI is cited as conserving an order of magnitude energy through optimized task allocation for general programs.Source: 2) Partial Offloading · p. 16
• DeepWear simplifies DL partial-offloading decisions by pruning computation-intensive nodes from a model DAG and grouping repeated sub-DAGs.The transformed structure enables lower-complexity decision making for allocating DL computation.Source: 2) Partial Offloading · p. 16
• IONN incrementally uploads model partitions, while unnecessary DNN layers are omitted to avoid additional offloading.Sequential model construction at the edge addresses delay from uploading an entire model at once.Source: 2) Partial Offloading · p. 16

Scalability

• End–edge offloading supports less computation-intensive services and small concurrent query volumes, but a single edge node is insufficient for many simultaneous queries.This limitation motivates collaboration across additional resources and architectural levels.Source: 3) Vertical Collaboration · p. 16

Vertical Collaboration

• Vertical collaboration profiles DNN layers and selects end, edge, or cloud partition points using latency, energy, wireless-condition, and server-load information.Neurosurgeon chooses the allocation that achieves the best latency or mobile-energy outcome under the modeled conditions.Source: 3) Vertical Collaboration · p. 16
• Vertical collaboration can absorb query floods through cloud assistance and reduce backbone traffic when edge preprocessing shrinks transmitted intermediate data, but it may add communication delay.Early exits can allow confident results to be produced before cloud processing, while cloud support handles workloads beyond edge capacity.Source: 3) Vertical Collaboration · p. 16

Tailoring Edge Frameworks for DL

• Customized edge frameworks match models to hardware, exploit performance and power characteristics, and orchestrate services, but early frameworks lack native offloading and collaboration support.Model deployment also requires architecture-aware mapping, automated porting, and online selection based on accuracy, latency, energy, and memory footprint.Source: C. Tailoring Edge Frameworks for DL · p. 17

Performance Evaluation for Edge DL

• Standardized evaluation of DL libraries on resource-constrained edge devices is needed to compare latency and memory performance.Impartial evaluation can guide software-stack optimization for specific edge hardware.Source: D. Performance Evaluation for Edge DL · p. 17
• A standard Edge DL testbed is missing, limiting evaluation of end-to-end architectures that combine edge computing, end devices, and the cloud.Proposed testbeds include openLEON and CAVBench, while integrated wireless, networking, service, and cloud models remain desirable.Source: D. Performance Evaluation for Edge DL · pp. 17–18

Distributed Training at Edge

• Communication-efficient distributed training compresses or sparsifies gradient and model updates, and online knowledge distillation exchanges outputs instead of full parameters.These approaches target the high communication cost of synchronizing large DL models.Source: A. Distributed Training at Edge · p. 18

Federated Learning

• Federated learning keeps raw data on edge devices while iteratively exchanging local model updates with an aggregation server.The survey presents FL as a way to address non-IID data, intermittent connectivity, unbalanced contributions, and privacy concerns.Source: B. Vanilla Federated Learning at Edge · p. 19
• FL still incurs communication costs for model updates, motivating periodic communication, subsampling, quantization, and other compression methods.Downlink and uplink updates can both be costly when models are large.Source: C. Communication-efficient FL · p. 19

Resource-optimized FL

• Heterogeneous FL devices can become stragglers, delaying aggregation and forcing trade-offs between model structure, resource use, and convergence.ELFISH dynamically masks neurons during training and improves training speed by 2×, but synchronous aggregation may not handle extreme situations.Source: D. Resource-optimized FL · p. 20
• FL wall-clock time combines client computation and communication, while energy consumption and training time can conflict.FEDL approximately solves each client’s local problem to balance energy cost and training time.Source: D. Resource-optimized FL · p. 20
• Joint resource allocation and user selection formulate FL optimization around minimizing loss under delay and energy constraints.The base station allocates bandwidth resources while users adjust transmit power to reduce packet errors.Source: D. Resource-optimized FL · p. 20
• Downlink reduction: 14×; Uplink reduction: 28×; Local computation reduction: 1.7×.These figures summarize reductions reported for a communication-efficient FL method.Source: D. Resource-optimized FL · p. 21
• Convergence acceleration: 62.4%.The reported result concerns convergence under the resource-optimized FL setting.Source: D. Resource-optimized FL · p. 21
• Top 1 accuracy improvement: 5.59%- 5.89%; Communication traffic reduction: 92%.The reported method jointly improves the listed accuracy measure and reduces communication traffic.Source: D. Resource-optimized FL · p. 21
• Reduction of the FL loss function value: up to 16%.This is the reported upper-bound reduction for the FL loss function.Source: D. Resource-optimized FL · p. 21

Security-enhanced FL

• Federated learning must tolerate poisoned training data, abnormal devices, unpredictable connectivity, and device dropouts.Secure Aggregation is described as tolerating up to one-third of devices failing to process or upload updates on time.Source: E. Security-enhanced FL · pp. 21–22
• BlockFL replaces dependence on a specific aggregation server with local global-model updating and adds rewards to stimulate edge-device participation.The approach targets server malfunction and unequal willingness to contribute.Source: E. Security-enhanced FL · p. 22

DL for Optimizing Edge

• Edge caching is difficult because content popularity varies spatiotemporally and heterogeneous, hierarchical networks complicate cache-policy design.Traditional methods also require many online optimization iterations.Source: A. DL for Adaptive Edge Caching · p. 22, 1) Use Cases of DNNs · p. 22
• DNNs can extract user and content features and use offline training to determine cache policies without repeating heavy online optimization.The learned input-solution relation depends on existing optimal or heuristic solutions.Source: 1) Use Cases of DNNs · p. 22
• DNN-based caching performance is bounded by the fixed optimization algorithms used to generate training solutions and is not self-adapted.This limitation applies to the surveyed DNN-based caching methods.Source: 1) Use Cases of DNNs · p. 22
• A DDPG agent with Wolpertinger architecture achieves competitive cache hit rates while reducing runtime in a single-base-station caching scenario.Wolpertinger narrows a large action space through a primary action set and k-nearest-neighbor mapping.Source: 2) Use Cases of DRL · p. 23

DL for Optimizing Edge Task Offloading

• Computation offloading is NP-hard because it must select receiving edge nodes, offloading ratios, and resource allocations under multiple constraints.Large state and action spaces make conventional learning algorithms difficult to apply directly.Source: B. DL for Optimizing Edge Task… · p. 23
• Offline optimal solutions can train a DNN to map the composite edge-network state to an offloading decision, avoiding online solution search.This transfers computation complexity from online optimization to DL training.Source: 1) Use Cases of DNNs · p. 23
• DQL can select target edge nodes and transmission modes online by accounting for communication conditions and offloading utility.The approach responds to potentially low-quality wireless environments.Source: 2) Use Cases of DRL · p. 23
• Double-DQL saves more energy and achieves higher training efficiency than DQL in the reported DVFS experiment.Preclassification can narrow the action space as the number of edge devices increases.Source: 2) Use Cases of DRL · pp. 23, 25
• CNN-based state compression and Q-function decomposition extend Double-DQL to more complex energy-harvesting offloading environments.Energy harvesting from ambient radio-frequency signals complicates the offloading problem.Source: 2) Use Cases of DRL · p. 25
• Using DRL only for the NP-hard offloading decision, while separating resource allocation, can reduce the action-space and computational burden.The surveyed design decomposes weighted-sum computation-rate maximization into offloading and resource-allocation subproblems.Source: 2) Use Cases of DRL · p. 25
• Offloading optimization is connected to broader edge management needs, including mobility-aware association, communication-mode control, and defense against wireless attacks.DL-based methods model interaction costs, control communication states, and support secure offloading.Source: C. DL for Edge Management and… · p. 25, 1) Edge Communication · p. 25, 2) Edge Security · p. 25

3) Joint Edge Optimization:

• Edge networks use DL and DRL for resource optimization tasks including routing, traffic control, caching, and computation offloading.Source: 3) Joint Edge Optimization · p. 26
• Multi-timescale DRL models vehicle mobility across time slots and epochs to address immediate mobility effects and large action spaces in IoV resource allocation.Source: 3) Joint Edge Optimization · p. 26
• Edge services require model-selection criteria beyond standard accuracy because wireless conditions and concurrent requests vary across edge networks.Source: 1) Ambiguous Performance Metrics · p. 26
• Customized models require coordination across end, edge, and cloud resources, including hybrid-precision designs that balance optimization against cloud-side accuracy risks.Source: 3) Hybrid model modification · p. 26
• Retraining optimized models can delay adaptation, while heterogeneous devices and dynamic networks create diverse customization requirements.Source: 4) Coordination between training and… · p. 27

C. Complete Edge Architecture for DL

• A complete DL edge framework spans data acquisition, service deployment, and task processing.Source: C. Complete Edge Architecture for DL · p. 27
• Edge architectures must efficiently acquire and process device-generated data before feeding it to DL models.Source: 1) Edge for Data Processing · p. 27
• Adaptive edge data acquisition and compression can reduce device workload, resource overhead, bandwidth pressure, and transmission delay, but heterogeneous DL data remain insufficiently addressed.Source: 1) Edge for Data Processing · p. 27
• Microservice-based edge deployment must isolate DL services on shared resources, but edge microservice frameworks remain immature.Source: 2) Microservice for Edge DL Services · p. 27
• Offloading heavy DL computation requires incentives for participating edge nodes and security against anonymous nodes.Source: 3) Incentive and trusty offloading… · p. 27
• Dividing computation-intensive DL models into subtasks enables task offloading and collaboration among end devices, edge nodes, base stations, and cloud infrastructure.Source: 4) Integration with “DL for optimizing… · p. 27
• Edge training is constrained by weak device performance and limited framework support, with many studies still simulating training theoretically.Source: D. Practical Training Principles at Edge · p. 27, 1) Data Parallelism versus Model… · pp. 27–28
• Federated learning can combine data and model parallelism, but practical edge training must handle unlabeled data and distributed resource constraints.Source: 1) Data Parallelism versus Model… · p. 28, 2) Where is training data from? · p. 28, Figure 21 · p. 28

3) Asynchronous FL at Edge:

• Synchronous federated learning scales only to hundreds of devices and may be inefficient under infrequent training and uncertain inter-device communication.Source: 3) Asynchronous FL at Edge · p. 28
• Device-specific scheduling and availability make epoch synchronization difficult, requiring server-selected participants and timeout rules for lagging devices.Source: 3) Asynchronous FL at Edge · p. 28
• Transfer learning and knowledge distillation can reduce edge training data, accelerate student convergence, improve accuracy, and limit overfitting.Source: 4) Transfer Learning-based Training · p. 28
• DL optimization spans applications such as content streaming, routing, and traffic control, but practical deployment must account for model overhead.Source: E. Deployment and Improvement of… · pp. 28–29
• Training-data transmission consumes bandwidth, while DL inference adds latency and can offset the benefits of edge optimization.Source: E. Deployment and Improvement of… · p. 29
• Open deployment questions concern model placement, architecture support, customized models, and training principles for timely edge management.Source: E. Deployment and Improvement of… · p. 29
• Multi-agent DRL and GNNs can represent end devices, edge nodes, and cloud systems as cooperating agents using local observations.Source: E. Deployment and Improvement of… · p. 29
• The survey frames Edge DL as a joint problem of optimizing training and inference under networking, communication, computing-power, and energy constraints.Source: §X · p. 29