Video super-resolution (VSR) aims to enhance low-resolution videos by leveraging both spatial and temporal information. While deep learning has led to impressive progress, it typically requires centralized data, which raises privacy concerns. Federated learning (FL) offers a privacy-friendly solution, but general FL frameworks often struggle with low-level vision tasks, resulting in blurry, low-quality outputs. To address this, we introduce FedVSR, the first FL framework specifically designed for VSR. It is architecture-agnostic and stateless, and introduces a lightweight loss function based on the Discrete Wavelet Transform (DWT) to better preserve high-frequency details during local training. Additionally, a loss-aware aggregation strategy combines both DWT-based and task-specific losses to guide global updates effectively. Extensive experiments across multiple VSR models and datasets show that FedVSR not only improves perceptual video quality (up to +0.89 dB PSNR, +0.0370 SSIM, -0.0347 LPIPS and 4.98 VMAF) but also achieves these gains with close to zero computation and communication overhead compared to its rivals. These results demonstrate Fed-VSR's potential to bridge the gap between privacy, efficiency, and perceptual quality, setting a new benchmark for federated learning in low-level vision tasks. Please refer to this link for the code. https://github.com/alimd94/FedVSR

Label correction methods are popular for their simple architecture in learning with noisy labels. However, they suffer severely from false label correction and achieve subpar performance compared with state-of-the-art methods. In this paper, we revisit the label correction methods through theoretical analysis of gradient scaling and demonstrate that the sample-wise dynamic and class-wise uniformity of interpolation weight prevents memorization of the mislabeled samples. We then propose DULC, a simple yet effective label correction method that uses the normalized Jensen-Shannon divergence (JSD) metric as the interpolation weight to promote sample-wise dynamic and class-wise uniformity. Additionally, we provide theoretical evidence that sharpening predictions in label correction facilitates the memorization of true class, and we achieve it by employing the augmentation strategy along with the sharpening function. Extensive experiments on CIFAR-10, CIFAR-100, TinyImageNet, WebVision and Clothing1M datasets demonstrate substantial improvements over state-of-the-art methods.

Adaptive optimizers such as Adam and RMSProp have gained attraction in complex neural networks, including generative adversarial networks (GANs) and Transformers, thanks to their stable performance and fast convergence compared to non-adaptive optimizers. A frequently overlooked limitation of adaptive optimizers is that adjusting the learning rate of each dimension individually would ignore the knowledge of the whole loss landscape, resulting in slow updates of parameters, invalidating the learning rate adjustment strategy and eventually leading to widespread insufficient convergence of parameters. In this paper, we propose HVAdam, a novel optimizer that associates all dimensions of the parameters to find a new parameter update direction, leading to a refined parameter update strategy for an increased convergence rate. We validated HVAdam in extensive experiments, showing its faster convergence, higher accuracy, and more stable performance on image classification, image generation, and natural language processing tasks. Particularly, HVAdam achieves a significant improvement on GANs compared with other state-of-the-art methods, especially in Wasserstein-GAN (WGAN) and its improved version with gradient penalty (WGAN-GP).

The top priority of a Housing and Homelessness System of Care (HHSC) is to connect people experiencing homelessness to supportive housing. An HHSC typically consists of many agencies serving the same population. Information technology platforms differ in type and quality between agencies, so their data are usually isolated from one agency to another. Larger agencies may have sufficient data to train and test artificial intelligence (AI) tools but smaller agencies typically do not. To address this gap, we introduce a Federated Learning (FL) approach enabling all agencies to train a predictive model collaboratively without sharing their sensitive data. We demonstrate how FL can be used within an HHSC to provide all agencies equitable access to quality AI and further assist human decision-makers in the allocation of resources within HHSC. This is achieved while preserving the privacy of the people within the data by not sharing identifying information between agencies without their consent. Our experimental results using real-world HHSC data from a North American city demonstrate that our FL approach offers comparable performance with the idealized scenario of training the predictive model with data fully shared and linked between agencies.

Federated Learning (FL) presents a paradigm shift towards distributed model training across isolated data repositories or edge devices without explicit data sharing. Despite of its advantages, FL is inherently less efficient than centralized training models, leading to increased energy consumption and, consequently, higher carbon emissions. In this paper, we propose CAMA, a carbon-aware FL framework, promoting the operation on renewable excess energy and spare computing capacity, aiming to minimize operational carbon emissions. CAMA introduces a dynamic model adaptation strategy which adapts the model sizes based on the availability of energy and computing resources. Ordered dropout is integratged to enable the aggregation with varying model sizes. Empirical evaluations on real-world energy and load traces demonstrate that our method achieves faster convergence and ensures equitable client participation, while scaling efficiently to handle large numbers of clients. The source code of CAMA is available at https://github.com/denoslab/CAMA.

Unmanned surface vessel (USV) has been widely used in various fields due to its autonomous advantages, and path planning is a crucial technology for autonomy. However, using global path planning alone cannot avoid moving obstacles, while using local path planning alone may lead to falling into local minima and fail to reach the target. Therefore, this article proposed the dynamic target artificial potential field (DTAPF) method which use a dynamic point that follows the global path generated by the A* algorithm as the target point of the artificial potential field (APF). In addition, in order to improve response time and safety of unmanned surface vessel (USV) navigation of the traditional centralized path planning methods, we proposed an edge computing architecture for global path planning and an offset guidance method to avoid moving obstacles while confirming to the collision regulations (CORLEGs). The experimental results show that, using the method proposed in this article, USV can reach the target in an environment with moving obstacles with high probability (about 99.4%), and compared to the traditional APF algorithm, our method can reduce collision probability by 71% with almost no increase in average path length and average navigation time. Besides, our architecture has much lower computing delay than local computing, and also lower than cloud computing.

Home exercise and self-served gyms allow a larger population to exercise regularly without the cost of hiring private coaches. In absence of professional guidance, however, exercisers can suffer from injuries to muscles and joints. High-precision, affordable arm tracking with commercial, off-the-shelf (COTS) wearable devices has become an urgent need to prevent workout injuries and improve exercise performance. Recent studies with inertial measurement units (IMUs) or audio signals are neither computationally feasible for real-time motion tracking with satisfactory accuracy using COTS devices nor practically usable due to the interference with noisy ambient environments. In this paper, we propose UltraMotion, a real-time, high-precision ultrasonic arm motion tracking system designed for practical use. UltraMotion performs point cloud queries based on hidden Markov models (HMMs), a novel ultrasonic acoustic ranging method, and an extended Kalman filter (EKF) to predict the locations of all three arm joints, making it the first system offering shoulder locations. Experimental results with only a smartphone and a smartwatch demonstrate the effectiveness of UltraMotion in tracking shoulder, elbow, and wrist locations with impressively small median errors of 6.4 cm, 7.1 cm, and 8.5 cm in real-world environments, outperforming all previous systems, making UltraMotion an ideal choice for daily exercise.

Deep Neural Networks (DNNs) possess powerful prediction capability thanks to their over-parameterization design, although the large model complexity makes it suffer from noisy supervision. Recent approaches seek to eliminate impacts from noisy labels by excluding data points with large loss values and showing promising performance. However, these approaches usually associate with significant computation overhead and lack of theoretical analysis. In this paper, we adopt a perspective to connect label noise with epistemic uncertainty. We design a simple, efficient, and theoretically provable robust algorithm named USDNL for DNNs with uncertainty-based Dropout. Specifically, we estimate the epistemic uncertainty of the network prediction after early training through single Dropout. The epistemic uncertainty is then combined with cross-entropy loss to select the clean samples during training. Finally, we theoretically show the equivalence of replacing selection loss with single cross-entropy loss. Compared to existing small-loss selection methods, USDNL features its simplicity for practical scenarios by only applying Dropout to a standard network, while still achieving high model accuracy. Extensive empirical results on both synthetic and real-world datasets show that USDNL outperforms other methods. Our code is available at https://github.com/kovelxyz/USDNL.

Federated learning (FL) sheds light on efficiently and privately learning from massive Internet of Things (IoT) devices. However, the iterative training and aggregation pose additional stress on the limited energy and availability budgets of clients. In this paper, we discuss two types of availability budgets of IoT clients, including the timing to start participating in FL and the communication budgets due to their constrained energy. We theoretically analyze the effect of availability budgets on FL, based on the availability constraints, by leveraging a decaying quadratic function to prioritize learning from statistically heterogeneous clients during the initial training rounds. We also consider the effects of client availability in terms of their participation to find a balance among clients with varying availability. We present FedCAB, an algorithm applying our theoretical model for the probabilistic rankings of the available clients to select in each round of FL model aggregation. Numerical results show the effectiveness of FedCAB under label distribution skew with a limited communication budget and clients that join the learning process in later rounds. We release the source code of FedCAB at https://github.com/denoslab/FedCAB.

Federated learning (FL) is a distributed and privacy-preserving learning framework for predictive modeling with massive data generated at the edge by Internet of Things (IoT) devices. One major challenge preventing the wide adoption of FL in IoT is the pervasive power supply constraints of IoT devices due to the intensive energy consumption of battery-powered clients for local training and model updates. Low battery levels of clients eventually lead to their early dropouts from edge networks, loss of training data jeopardizing the performance of FL, and their availability to perform other designated tasks. In this paper, we propose FedLE, an energy-efficient client selection framework that enables lifespan extension of edge IoT networks. In FedLE, the clients first run for a minimum epoch to generate their local model update. The models are partially uploaded to the server for calculating similarities between each pair of clients. Clustering is performed against these client pairs to identify those with similar model distributions. In each round, low-powered clients have a lower probability of being selected, delaying the draining of their batteries. Empirical studies show that FedLE outperforms baselines on benchmark datasets and lasts more training rounds than FedAvg with battery power constraints.

The rise of Federated Learning (FL) is bringing machine learning to edge computing by utilizing data scattered across edge devices. However, the heterogeneity of edge network topologies and the uncertainty of wireless transmission are two major obstructions of FL’s wide application in edge computing, leading to prohibitive convergence time and high communication cost. In this work, we propose an FL scheme to address both challenges simultaneously. Specifically, we enable edge devices to learn self-distilled neural networks that are readily prunable to arbitrary sizes, which capture the knowledge of the learning domain in a nested and progressive manner. Not only does our approach tackle system heterogeneity by serving edge devices with varying model architectures, but it also alleviates the issue of connection uncertainty by allowing transmitting part of the model parameters under faulty network connections, without wasting the contributing knowledge of the transmitted parameters. Extensive empirical studies show that under system heterogeneity and network instability, our approach demonstrates significant resilience and higher communication efficiency compared to the state-of-the-art.

Recent analytical research has pointed out that almost all vehicles spend over 95% of their time in parking lots where their powerful computing resources are wasted. In this paper, we propose a novel collaborative computing paradigm that efficiently offloads online heterogeneous computation tasks to parked vehicles (PVs) during peak hours. A container orchestration based on Kubernetes is advocated to integrate into the existing infrastructure due to its cutting-edge features such as auto-healing, load-balancing, and security. We formulate the offloading problem analytically and present an intelligent metaheuristic algorithm to address dynamic online demands. Extensive evaluation demonstrates that our proposed paradigm improves task arrival ratio and average offloading cost for more than 40% compared with a set of heuristic algorithms. Additionally, owners of PVs can be beneficial by sharing their idle vehicle resources through received incentives.

For embracing the ubiquitous Internet‐of‐Things (IoT) devices, edge computing and Network Function Virtualization (NFV) have been enabled in branch offices and homes in the form of virtual Customer‐Premises Equipment (vCPE). A Service Provider (SP) deploys vCPE instances as Virtual Network Functions (VNFs) on top of generic physical Customer‐Premises Equipment (pCPE) to ease administration. Upon a usage surge of IoT devices at a certain part of the network, vCPU, memory, and other resource limitations of a single pCPE node make it difficult to add new services handling the high demand. In this paper, we present IoT‐B&B, a novel architecture featuring resource sharing of pCPE nodes. When a pCPE node has sharable resources available, the SP will utilize its free resources as a “bed‐and‐breakfast” place to deploy vCPE instances in need. A placement algorithm is also presented to assign vCPE instances to a cost‐efficient pCPE node. By keeping vCPE instances at the network edge, their costs of hosting are reduced. Meanwhile, the transmission latencies are maintained at acceptable levels for processing real‐time data burst from IoT devices. The traffic load to the remote, centralized cloud can be substantially reduced.

Mobile edge computing (MEC) literally pushes cloud computing from remote datacenters to the life radius of end users. By leveraging the widely adopted ETSI network function virtualization (NFV) architecture, MEC provisions elastic and resilient mobile edge applications with proximity. Typical MEC virtualization infrastructure allows configurable placement policy to deploy mobile edge applications as virtual machines (VMs): affinity can be used to put VMs on the same host for inter-VM networking performance, while anti-affinity is to separate VMs for high availability. In this paper, we propose a novel model to track the availability and cost impact from placement policy changes of the mobile edge applications. We formulate our model as a stochastic programming problem. To minimize complexity challenge, we also propose a heuristic algorithm. With our model, the unit resource cost increases when there are less resources left on a host. Applying affinity would take up more resources of the host but saves network bandwidth cost because of co-location. When enforcing anti-affinity, experimental results show increases of both availability and inter-host network bandwidth cost. For applications with different resource requirements, our model is able to find their sweet points with the consideration of both resource cost and application availability, which is vital in a less robust MEC cloud environment.

For embracing the ubiquitous Internet-of-Things (IoT) devices, edge computing and Network Function Virtualization (NFV) have been enabled in branch offices and homes to provide network functions on top of generic physical Customer-Premises Equipment (pCPE). While latency can be greatly reduced as most traffic does not need to be transmitted to a remote, centralized cloud, the resource limitation of a single pCPE makes it difficult for VNFs to be elastic enough upon usage surge. In this paper, we present VNF-B&B, an architecture featuring resource sharing of pCPE across the network edge. SP utilizes idle, shareable pCPE nodes as bed-and-breakfast places to deploy VNFs of other users for a certain period. By keeping the VNFs at the network edge, the cost is minimized for processing real-time data burst from IoT devices. Meanwhile, the traffic load to the core network and service delay is substantially reduced.

Cloud computing and Network Function Virtualization (NFV) have been leveraged in building Internet-of-Things (IoT) networks because of their ability to reduce cost and to provide elastic and resilient services to rapid-growing IoT devices. In this paper, we propose a novel model to track the impact from placement strategy changes to the cost and the availability of deploying VMs of a Virtual Network Functions (VNF). With our model, the unit resource cost increases as there is less resource left on a host. Putting VMs on the same host takes up more resources of the host but saves network bandwidth cost because of co-location. When enforcing anti-affinity for placing VMs on different hosts, experimental results show increases of both availability and inter-host network bandwidth cost, but varying trends of total costs considering other resources.

The energy consumption rates of nodes vary in different positions in sensor networks, leading to the early depletion of the nodes and the disconnection of a network. In this paper, we present a distributed mobile sink guiding scheme (DLING) based on the center of gravity theory in physics. DLING regards the sensor network as an object composed of particles. Sensor nodes act as particles and their residual energy is treated as their "weight". We use the traditional discovering method for the center of gravity to compute the "energy center" of a network. The nodes near the energy center have more residual energy and can afford more traffic load. Then we propose a novel scheme to guide the sink sojourning around the network's center of energy so that the energy consumption of the nodes will be balanced. The simulation results demonstrate that DLING can achieve further improvements on the network lifetime in comparison with previous moving schemes while having low control costs.

Nowadays, the harmful blue-green algae blooms on lakes or streams threaten the daily life of millions of people in China. In this paper, we demonstrate the sensor network system we built on Lake Tai for the surveillance and cleanup of the algae blooms which is at work in Wuxi City, Jiangsu Province. We designed the sensor device and algorithm to monitor the algae bloom and estimate the bloom area. When the bloom area goes beyond the threshold, the salvaging boats are automatically dispatched to the scene to clean up the bloom for recycling utilization according to a mechanism in a scalable coordination fashion. The dispatching mechanism includes to consider the locations of salvaging boats, the number of available salvaging boats and the facility status in eight algae harvesting factories around the lake. The system also balances workloads of factories to achieve an overall high working efficiency. We develop a GIS-based management website for the end user to monitor the running of the whole system. All the sensing stations, the real time performance of the algae harvesting factory, the locations of salvaging boats and the automatically generated cleanup schedule are displayed in the map.