Deploying Convolutional Neural Network (CNN) models on ubiquitous Internet of Things (IoT) devices in a cloud-assisted manner to provide users with a variety of high-quality services has become mainstream. Most existing studies speed up model cloud training/on-device inference by reducing the number of convolution (Conv) parameters and floating-point operations (FLOPs). However, they usually employ two or more lightweight operations (e.g., depthwise Conv,$1\times 1$cheap Conv) to replace a Conv, which can still affect the model's speedup even with fewer parameters and FLOPs. To this end, we propose the Grouped NonLinear transformation generation method (GroupNL), leveraging data-agnostic, hyperparameters-fixed, and lightweight Nonlinear Transformation Functions (NLFs) to generate diversified feature maps on demand via grouping, thereby reducing resource consumption while improving the robustness of CNNs. First, in a GroupNL Conv layer, a small set of feature maps, i.e., seed feature maps, are generated based on the seed Conv operation. Then, we split seed feature maps into several groups, each with a set of different NLFs, to generate the required number of diversified feature maps with tensor manipulation operators and nonlinear processing in a lightweight manner without additional Conv operations. We further introduce a sparse GroupNL Conv to speed up by reasonably designing the seed Conv groups between the number of input channels and seed feature maps. Experiments conducted on benchmarks and on-device resource measurements demonstrate that the GroupNL Conv is an impressive alternative to Conv layers in baseline models. Specifically, on Icons-50 dataset, the accuracy of GroupNL-ResNet-18 is 2.86% higher than ResNet-18; on ImageNet-C dataset, the accuracy of GroupNL-EfficientNet-ES achieves about 1.1% higher than EfficientNet-ES. In addition, we verified the efficiency of GroupNL-based models in terms of cloud training and on-device inference.

Using low-rank adaptation to fine-tune pre-trained neural network models has attracted widespread attention due to its advantages of low resource requirements, high precision, and no additional inference delay. However, most existing methods are designed for large language models based on the Transformer structure and lack adaptation to convolutional neural networks (CNN) widely used on Internet of Things (IoT) devices. In addition, IoT devices are usually deployed outdoors and collect large amounts of data that are affected by environmental conditions. Providing a highly robust CNN model is a prerequisite for providing high-quality services. To this end, this paper proposes a highly robust seed-filter adaptation method (SeFA) for pre-trained CNNs. SeFA introduces an adaptation branch with the same structure as the backbone network. In the adaptation branch, some filters are first designated seed filters and grouped. Then, additional filters are generated from the grouped seed filters and nonlinear transformation functions (NLFs) with different hyperparameters. The seed filters' parameters are updated during model training, and the NLF hyperparameters are randomly initialized and frozen. Both grouping seed filters and configuring NLFs with non-learnable hyperparameters can improve the model's robustness. This is because grouping seed filters can generate diverse filters on demand without increasing the model's complexity, and the NLFs' rules can regularize the model. It is worth mentioning that the number of fine-tuning parameters of the pre-trained model that can adapt to downstream tasks can be flexibly controlled by specifying the number of seed filters. The key idea of this paper is to propose SeFA with flexible, controllable, learnable parameters, high robustness, and adaptability to pretrained CNN models, thereby facilitating fine-tuning on resource-constrained IoT devices and providing highly robust visual services. Experimental results on the CIFAR-10, CIFAR-10-C, CIFAR-100, CIFAR-100-C, and Icons-50 datasets demonstrate that the proposed SeFA outperforms other state-of-the-art methods. Specifically, on the ResNet-152 model and the CIFAR-10-C dataset, the accuracy of our SeFA is about 7% higher than that of the full fine-tuning method. In addition, we verify the efficiency of SeFA-based models for fine-tuning and inference on the device.

Cloud-assisted resource-constrained and widely used outdoor IoT devices deploying highly robust convolutional neural network (CNN) models to provide high-quality visual services have attracted widespread attention from industry and academia. Most existing methods suffer from two limitations: (i) large amount of parameter transmission, and (ii) low robustness in handling data affected by the environment. To this end, this paper proposes a transmission-friendly and robust CNN design method called NonLinear transformation generation Filter with NonLearnable hyperparameters, namely NL2Filter. In NL2Filter, some filters are first designated as seed filters, whose parameters are learnable, that is, updated as the model is trained. Other filters in this layer are generated based on the seed filters and nonlinear transformation function (NLF). The hyperparameters of the NLF in NL2Filter are randomly initialized and remain unchanged, so they can be saved and reproduced using random seed. After the cloud server trains the NL2Filter-CNN model, it only needs to send a small number of learnable parameters and random seed to reproduce the CNN model trained by the cloud server. Compared with sending the complete CNN model, sending only a small number of learnable parameters and random seed can significantly reduce the number of model parameters sent by the cloud server. On the other hand, NL2Filter groups seed filters and then uses NLFs with different hyperparameters to generate diverse filters on demand for each group, thereby improving the model's ability to capture feature diversity without increasing the complexity of the model, thereby improving the model's robustness in processing data affected by the environment. Experimental results on CIFAR-10, CIFAR-10-C, and Icons-50 datasets demonstrate that the proposed NL2Filter outperforms other state-of-the-art methods. Specifically, using the ResNet-101 architecture, on the CIFAR-10-C, NL2Filter's accuracy is about 3.9% higher than that of MonoCNN; On the Icons-50, NL2Filter's accuracy is about 2.7% higher than that of MonoCNN and about 2.3% higher than that of the standard ResNet-101.

This study aims to guide pre-service teachers in effectively utilizing Generative Artificial Intelligence (GAI) tools to enhance their Artificial Intelligence (AI) literacy. By ana-lyzing the impact of GAI on the AI literacy of pre-service teachers, this study proposes a “teacher-student-machine” triadic interactive teaching model based on self-organized learning and deep empowerment through GAI. Using an experimental class in a software engineering course as a case study, we construct an end-to-end experimental environment for the project-driven teaching lifecycle, covering requirement analysis, system design, development and implementation, as well as software testing and maintenance phases. GAI is used to support the completion of experimental tasks while fostering pre-service teachers' AI literacy in five core competencies: data literacy, digital communication and collaboration, critical thinking, computational thinking, and ethical literacy. The study demonstrates that this model effectively improves the AI literacy of pre-service teachers, provides a reference for innovating software engineering practice teaching, and offers a useful framework for the application of GAI in education.

This study investigates the trade-off between system stability and offloading cost in collaborative edge computing. While collaborative offloading among multiple edge servers enhances resource utilization, existing methods often overlook the role of queue stability in overall system performance. To address this, a multi-hop data transmission model is developed, along with a cost model that captures both energy consumption and delay. A time-varying queue model is then introduced to maintain system stability. Based on Lyapunov optimization, a dynamic offloading algorithm (LDSO) is proposed to minimize offloading cost while ensuring long-term stability. Theoretical analysis and experimental results verify that the proposed LDSO achieves significant improvements in both cost efficiency and system stability compared to the state-of-the-art.

Using low-rank adaptation to fine-tuning pretrained neural network models has attracted widespread attention due to its advantages of low resource requirements, high precision, and no additional inference delay. However, most existing methods are designed for large language models based on Transformer structure and lack adaptation to convolutional neural networks (CNN) widely used on Internet of Things (IoT) devices. In addition, IoT devices are usually deployed outdoors and collect a large amount of data affected by the environment. Providing a highly robust model is a prerequisite for providing high-quality services. To this end, this paper proposes a highly robust seed-filter adaptation method (SeFA) for pre-trained CNNs. SeFA introduces an adaptation branch with the same structure as the backbone network. In the adaptation branch, some filters are first designated seed filters and grouped. Then, other filters are generated based on the grouped seed filters and nonlinear transformation functions (NLFs) with different hyperparameters. The parameters of the seed filters are updated with model training, and the hyperparameters of the NLFs are randomly initialized and frozen. Both grouping seed filters and configuring NLFs with nonlearnable hyperparameters can improve the robustness of the model. This is because grouping seed filters can generate diverse filters on demand without increasing the model's complexity, and the NLFs' rules can regularize the model. The key idea of this paper is to propose SeFA with flexible controllable learnable parameters, high robustness and adaptability to pretrained CNN models, to facilitate fine-tuning of pre-trained CNN models on resource-constrained IoT devices to provide highly robust visual services. Experimental results on the CIFAR-10, CIFAR-10-C, CIFAR-100, CIFAR-100-C, and Icons50 datasets demonstrate that the proposed SeFA outperforms other state-of-the-art methods. Specifically, based on the ResNet152, on the CIFAR-10-C dataset, the accuracy of our SeFA is about$+7 {\%}$higher than that of the full fine-tuning method.

Using pruning techniques to prune redundant weights in large language models (LLMs) for model size reduction, enabling deployment on devices to deliver high-quality services, has garnered significant attention from industry and academia. However, most existing pruning methods suffer from two major problems: (1) they rely on operations with high computational complexity, such as Hessian matrix calculation, and (2) high pruning rate leads to a significant decrease in model accuracy. To this end, this paper proposes a low resource requirement and low accuracy loss post-training pruning approach, namely the compensated post-training pruning method (CPP) for on-device LLM services. First, CPP employs singular value decomposition on weight matrices, sorting the decomposed singular values in descending order. Based on the pruning ratio, it retains the principal eigenvector corresponding to the larger singular values. To maintain consistent output feature distributions, CPP applies orthogonal transformations to input data and weight matrices, leveraging the principle of matrix orthogonality invariance. Compared with pruning approaches based on the Hessian matrix, CPP does not need to iteratively calculate the second derivative, thereby avoiding a high computational overhead. Second, the CPP incorporates bias compensation to use valuable information in pruned weights to improve the accuracy of the model. It constructs mapping relationships between input features and pruned weights through tensor decomposition techniques to generate bias compensation terms. These terms fine-tune the output of each layer, reducing pruning-induced errors from ratio-based pruning, and consequently improving model accuracy. Finally, CPP is validated through comprehensive experiments on nine benchmark datasets (e.g., WikiText-2 and PIQA) using large language models (i.e., LLaMA and OPT). Specifically, compared with the FLAP, CPP demonstrates 6.68% reduction in average pruning time. At pruning ratios of 10% and 20%, CPP achieves 2. 14% and 1.45% reductions in perplexity, respectively, along with 1.22% to 3.43% improvement in zeroshot inference accuracy.

Task offloading can meet users’ demands for the latency and energy consumption by offloading tasks from resource-constrained Internet of Things devices to relatively resource-rich edge servers. Traditional task offloading usually makes use of fixed base stations or servers as edge servers. This would lead to limited range of services and increased costs due to large-scale deployment of edge servers. Therefore, deploying unmanned aerial vehicles (UAVs) as mobile edge servers for task offloading in complex terrains (e.g., forest, desert, etc.) is a worthwhile research problem. To this end, this article proposes a UAV-assisted task offloading mechanism. The mechanism aims to minimize the weighted sum of latency and energy consumption through jointly optimizing resource allocation, offloading decision, and UAV trajectory. We first transform the nonconvex optimization problem into convex optimization subproblems to obtain the optimal resource allocation. Second, we use an improved particle swarm optimization algorithm to find the optimal offloading decision. Finally, we present the deep determination policy gradient algorithm to optimize the UAV trajectory which is a kind of deep reinforcement learning algorithm. Through simulation experiments, we show that the proposed mechanism can efficiently reduce the weighted sum of latency and energy consumption.

Vehicle Edge Computing (VEC) can effectively improve the efficiency of vehicle task calculation and offloading by integrating edge computing and the Internet of Vehicles. However, current research in VEC mostly focuses on device collaboration within single-hop or two-hop ranges, limiting the additional performance gain and load balancing provided by multi-hop device collaboration. In this study, an Edge Digital Twin assisted Multi-hop Task Offloading Scheme (EDT_MTOS) is proposed to enhance the execution efficiency of vehicle tasks by establishing an edge digital twin layer for virtual mapping of vehicles and edge servers. Firstly, the multi-hop task offloading problem is transformed into a cost optimization problem related to delay and energy consumption under the constraint of load balancing. Secondly, a dynamic collaboration knowledge graph based on digital twin knowledge mapping is introduced to select collaborative device sets for the upload and return links within the multi-hop range. Then, a value iteration algorithm based on the maximum link quality is proposed to realize the dynamic collaboration knowledge graph. Furthermore, a task offloading solution algorithm is proposed based on Dynamic collaboration Knowledge Graph and Double Deep Q-Network (DKG_DDQN). Finally, the simulation results demonstrate that the proposed offloading algorithm can reduce vehicle task processing costs by 33.77% and 26.53% in an idle scenario, and by 27.60% and 26.21% in a busy scenario, compared to state-of-the-art collaborative algorithms such as COOR and DRL-COMV, respectively.

In cell-free massive multiple-input multiple-output (CF-mMIMO) systems, seamless communication coverage is achieved through the dense deployment of numerous access points (APs), significantly enhancing spectral efficiency (SE) for users and overall system capacity. However, the implementation of this approach demands substantial deployment costs and unavoidably necessitates the use of non-ideal hardware. This paper investigates the achievable rate of users in the uplink CF-mMIMO systems that employ nonlinear power amplifiers (PAs) and low-resolution analog-to-digital converters (ADCs) at user equipment (UE) and APs, respectively. In particular, we derive a closed-form expression for the achievable uplink user rate and conduct a comprehensive analysis of various factors, including the number of APs, UE density, number of AP antennas, and ADC resolution. To mitigate the interference among UEs and maximize the sum rate, we propose a graph neural network (GNN) assisted actor-critic algorithm (DMAGNN-AC) for power allocation. The established framework overcomes the representation bottleneck of DRL in high-dimensional unstructured state spaces and provides physically interpretable feature embeddings. In comparison to the full power output, the proposed power allocation scheme is capable of doubling the rate. Furthermore, to address the detrimental impact of low-resolution ADCs on the rate, we develop an enhanced algorithm, multi-agent deep Q-integrated network (MADQIN), which optimizes the allocation strategy of ADC resolutions. Finally, the effectiveness of the proposed schemes is validated by the presented simulation results.

As one of the most critical tasks currently, it is essential to excavate the status update performance. In this paper, we combine access control and pilot design to minimize average age of information (AoI) in cell-free (CF) massive multiple-input multiple-output (MIMO) system with grant-free (GF) random access, where the positions of both the access points (APs) and machine-type communication devices (MTCDs) are geographically distributed as independent Poisson point processes (PPPs). Based on the establish two-disk computation model, we frist formulate the access successful probability by leveraging the stochastic geometry for GF transmission. Then, the closed-form expression of the average AoI is approximately derived. The derived results enable us precisely quantify the impact of network parameters on the system-level performance. To minimize the average AoI, a joint access control and pilot design scheme is proposed. Finally, simulation results are provided to furnish invaluable insight into system performance and certificate the validity of the proposed algorithm.

As the size of data clusters grows, the issues of storage redundancy, data availability, and the cost of edge caching systems become the focus of attention. Although the multi-copy strategy improves data availability, its redundancy cost is high and does not fully consider file data characteristics. Erasure codes techniques have been widely studied as an alternative, but the selection of a suitable coding scheme remains a challenge. To this end, we propose a hybrid coding caching scheme. Firstly, we predict data types using machine learning methods to ensure that the most valuable content is stored. Secondly, for the characteristics of the multi-copy strategy, low-density parity check (LDPC), and Reed-Solomon coding (RS), we propose an adaptive data partitioning strategy (ADPRF) to select appropriate coding schemes for different data. Finally, by combining the ideas of dynamic programming and heuristic algorithms, we propose a redundancy-aware edge collaborative caching algorithm (RPCO) with joint optimization of energy and cost to determine the optimal set of collaborative nodes and the optimal cache locations for files. Compared with the traditional responsive caching scheme, this algorithm can reduce the system cost by 13.8% and access latency by 11.5%.

When offloading computing tasks of sensory data to the edge network, it is necessary to consider whether the idle resources such as CPU frequency and memory, meet the task processing requirements. However, even if edge collaboration is used to improve offloading performance, most studies assume homogeneity in hardware configuration across all edge servers, discarding the impact of the differentiated resource allocation among heterogeneous edge servers. Therefore, resource allocation and offloading decisions in a collaborative heterogeneous edge network are comprehensively considered in this study. Firstly, the offloading problem of heterogeneous edge servers is expressed as a joint optimization problem associated with delay and energy consumption constrained by CPU frequency and storage resources. Secondly, dynamic collaboration clusters are constructed based on distance, position and workload correlation to identify distinct collaboration regions and balance the load within edge servers. And then, a distributed alternating direction multiplier method (ADMM) based on constraint projection and variable splitting is proposed to solve the optimization problem. Additionally, a cooperative path selection algorithm, which takes into account length and throughput of return paths, is proposed to alleviate network congestion and minimize energy consumption loss. Finally, the proposed algorithm for Collaborative offloading and resource Configuration (CooCo) is demonstrated to be effective and rapidly converging based on a real dataset from Shanghai Telecom. The simulation results also show that compared to the DRAOA, no-cooperation, single-hop, and other state-of-the-art collaborative algorithm, CooCo can significantly reduce the sum of the system costs by 26%, 35%,11% and 8%, respectively.

Mobile devices commonly offload latency-sensitive applications to edge servers to meet low-latency requirements. However, existing studies overlook dependency and application hit ratio considerations, hindering effective offloading for multi-applications and multi-tasks. To this end, this article proposes a Dependent task offloading and Service placement Optimization (DSO) method to maximize the application hit ratio, thereby providing high-quality service. The proposed DSO includes Improved Multi-Agent Q-Learning (IMAQL) and greedy algorithms. IMAQL optimizes service placement via Q-learning, while the greedy algorithm schedules task offloading. Extensive experiments on public datasets demonstrate that the DSO method enhances the application hit ratio by 4.7% to 11.7% and reduces the completion time by about 3.4% to 4.9% compared to alternative approaches.

It has become mainstream for mobile devices to offload latency-sensitive applications to edge servers for execution to meet low-latency requirements. However, the existing related studies lack the consideration of application hit ratio, which makes them unable to meet the increasingly complex offloading of multi-applications including multi-tasks. To this end, this paper proposes a Multi-task offloading and Service placement optimization (MSO) method with the goal of maximizing the application hit ratio to provide high-quality service. The proposed MSO is constructed with Improved Multi-Agent Q-Learning (IMAQL) and load-balancing algorithms. IMAQL aims to learn an optimal service placement policy by using Q-learning techniques. Next, the load-balancing algorithm is designed to offload tasks according to the service placement policy. To verify the effectiveness of the MSO method, we conduct extensive experiments on a publicly available dataset. The experimental results show that the proposed method can improve the application hit ratio by appropriately 2.6% to 9% compared with other methods.

Task offloading offloads latency-sensitive and computation-intensive applications from resource-constrained terminal devices to relatively resource-rich edge servers to meet users’ demands for latency and energy consumption, which has attracted extensive attention from academia and industry. However, most of the existing researches only considers offloading dependent tasks within a single application or multiple independent applications, while ignoring the dependencies between applications. To this end, this paper proposes an offloading strategy for distributed dependent applications under the condition of limited computing and cache resources. The goal of the proposed strategy is to minimize the weighted sum of latency and energy to complete all applications while solving the offloading and resource allocation problems of dependent applications. However, the dual dependencies between applications and tasks within the application complicate offloading tasks. To accommodate this issue, we represent the dual dependencies as a directed acyclic graph. Then, we design the offloading strategy as follows: First, we transform the formulated non-convex problem into convex optimization subproblems. Second, we iteratively calculate the task priority and obtain the optimal offloading decision of the task according to the priority. Finally, we perform validation on real datasets. Compared with several state-of-the-art methods, our proposed strategy can significantly reduce the weighted sum of latency and energy.

With the easy access to mobile networks and the proliferation of video applications, video traffic is occupying a great portion of the network traffic, which poses a new challenge of how to alleviate the heavy backhaul traffic and ensure the high quality of experience for video services. As a promising solution towards addressing this challenge, video caching in edge networks has recently received significant attention, which mostly considers the video popularity and the user preference for the video. However, few studies consider the user behavior and the user preference for different parts of the video that indeed have an essential impact on caching efficiency. Hence, this article proposes a new caching and resource scheduling scheme for adaptive bitrate videos by incorporating these fine-grained factors. We first model the video service problem as a nonlinear integer programming problem, which can be divided into a cache placement problem and an online resource scheduling problem. Then, we design efficient algorithms based on several techniques, including greedy strategy, relaxation, and rounding, to solve the two problems. Extensive experimental results based on two real-world datasets show that the proposed solution achieves superior performance compared with several state-of-the-art caching approaches.

There have been many studies focusing on edge server deployment and service placement in mobile-edge computing (MEC), respectively, but rare works took both of them into consideration. However, edge server deployment and service placement are coupling issues in practice, where the former affects the latter. Besides, the economic benefit of the MEC platform is also a consideration. Due to different service request rates and prices, appropriate service placement solutions are needed to increase the overall profit. In this article, we propose a complete process combining edge server and service placement, where service placement explicitly takes into account the structure of current edge server placement and different service request rates and prices. We design a joint edge server deployment and service placement model with the goal of maximizing the overall profit of all edge servers under the constraints of the number of edge servers, the relationship among edge servers and base stations, the storage capacity, and the computing capacity of each edge server. We propose a two-step method including the clustering algorithm and nonlinear programming to solve the formulated problem. Extensive evaluations based on the real-world data set demonstrate that the proposed algorithm outperforms the baseline methods.

Spatial crowdsourcing (SC) is a popular distributed problem-solving paradigm that harnesses the power of mobile workers (e.g., smartphone users) to perform location-based tasks (e.g., checking product placement or taking landmark photos). Typically, a worker needs to travel physically to the target location to finish the assigned task. Hence, the worker’s familiarity level on the target location directly influences the completion quality of the task. In addition, from the perspective of the SC server, it is desirable to finish all tasks with a low recruitment cost. Combining these issues, we propose a Bi-Objective Task Planning (BOTP) problem in SC, where the server makes a task assignment and schedule for the workers to jointly optimize the workers’ familiarity levels on the locations of assigned tasks and the total cost of worker recruitment. The BOTP problem is proved to be NP-hard and thus intractable. To solve this challenging problem, we propose two algorithms: a divide-and-conquer algorithm based on the constraint method and a heuristic algorithm based on the multi-objective simulated annealing algorithm. The extensive evaluations on a real-world dataset demonstrate the effectiveness of the proposed algorithms.

Mobile Edge Computing (MEC) can provide the user with computation resources close to the user. In order to cope with stringent requirements of applications on latency and reduce energy consumption, MEC enables to offload highly computation demanding tasks to the edge servers. In this paper, we present a brief overview of user-oriented computation offloading in MEC. We first describe the definition of computation offloading decision, quality of service metric and research scenario. Then we review the existing user-oriented computation offloading approaches from the single user scenario and multiple user scenarios. Finally, we conclude the paper.

In order to build highly reliable composite service via Service Oriented Architecture (SOA) in the Mobile Fog Computing environment, various fault tolerance strategies have been widely studied and got notable achievements. In this paper, we provide a comprehensive overview of key fault tolerance strategies. Firstly, fault tolerance strategies are categorized into static and dynamic fault tolerance according to the phase of their adoption. Secondly, we review various static fault tolerance strategies. Then, dynamic fault tolerance implementation mechanisms are analyzed. Finally, main challenges confronted by fault tolerance for composite service are reviewed.

Given a service composition, it is challenging but important to have a runtime adaptation, due to the complicated execution environment and evolving feature of Web service. In this paper, we present a runtime adaptive service composition approach, taking execution time minimization and revenue maximization into consideration. Based on dynamic programming, we deduce the optimal policy. Through this policy, orchestrator selects one concrete service for per task on runtime. The experimental results show that the proposed approach outperforms previous approach.