Semi-supervised learning (SSL) enhances model generalizability by jointly leveraging labeled and unlabeled data. Nevertheless, the closed-world assumption of SSL always fails in open-world scenarios, where unlabeled data often contains novel classes. To address this limitation, open-world SSL (OWSSL) has been proposed as a more realistic paradigm, aiming not only to recognize known classes but also to discover novel classes. Existing OWSSL methods typically rely on representation similarity and pseudo-labeling to discriminate classes. However, during model training, these methods neglect the inherent class-prediction bias, consequently leading to self-reinforcing confirmation bias in pseudo-labels and representation confusion for hard novel classes. To address these critical challenges, we propose a Open-world Bias-guided Pseudo-label Refinement approach, named OpenBPR, which is the first to regard class prediction bias as the reference to guide the debiased pseudo-labeling and class representation decoupling. In OpenBPR, we propose a debiased pseudo-labeling method based on expectation-maximization, which exploits class prediction bias to dynamically optimize pseudo-labels, effectively alleviating confirmation bias in pseudo-labels. Furthermore, we propose a class-aware representation decoupling strategy for hard novel classes, which decouples representations by the designed competitive class decoupling regularization to assist in improving the refinement performance of pseudo-labels. Experimental results on a series of benchmark datasets demonstrate that OpenBPR outperforms state-of-the-art methods in discriminating both known and novel classes.

Classifiers often learn to be biased corresponding to the class-imbalanced dataset under the semi-supervised learning (SSL) set. While previous work tries to appropriately re-balance the classifiers by subtracting a class-irrelevant image's logit, we further utilize a cheaper form of consistency gradients, which can be widely applicable to various class-imbalanced SSL (CISSL) models. We theoretically analyze that the process of refining pseudo-labels with a baseline image (solid color image without any patterns) in the basic SSL algorithm implicitly utilizes integrated gradient flow training, which can improve the attribution ability. Based on the analysis, we propose a consistently conflicting gradient-based debiasing scheme dubbed LCGC, by encouraging biased class predictions during training. We intentionally update the pseudo-labels whose gradient conflicts with the debiased logits, which is represented as the optimization direction offered by the over-imbalanced classifier predictions. Then, we debias the predictions by subtraction the baseline image logits during testing. Extensive experiments demonstrate that our method can significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Advances in prompt engineering and fine-tuning techniques have further enhanced their ability to address complex reasoning challenges. However, these advanced capabilities are often exclusive to models exceeding 100 billion parameters. Although Chain-of-Thought (CoT) fine-tuning methods have been explored for smaller models (under 10 billion parameters), they typically depend on extensive CoT training data, which can introduce inconsistencies and limit effectiveness in low-data settings. To overcome these limitations, this paper introduce a new reasoning strategy Solution Guidance (SG) and a plug-and-play training paradigm Solution-Guidance Fine-Tuning (SGFT) for enhancing the reasoning capabilities of small language models. SG focuses on problem understanding and decomposition at the semantic and logical levels, rather than specific computations, which can effectively improve the SLMs’ generalization and reasoning abilities. With only a small amount of SG training data, SGFT can fine-tune a SLM to produce accurate problem-solving guidances, which can then be flexibly fed to any SLM as prompts, enabling it to generate correct answers directly. Experimental results demonstrate that our method significantly improves the performance of SLMs on various reasoning tasks, enhancing both their practicality and efficiency within resource-constrained environments.

Understanding the dynamics of causal structures is crucial for uncovering the underlying processes in time series data. Previous approaches rely on static assumptions, where contemporaneous and time-lagged dependencies are assumed to have invariant topological structures. However, these models fail to capture the evolving causal relationship between variables when the underlying process exhibits such dynamics. To address this limitation, we propose DyCAST, a novel framework designed to learn dynamic causal structures in time series using Neural Ordinary Differential Equations (Neural ODEs). The key innovation lies in modeling the temporal dynamics of the contemporaneous structure, drawing inspiration from recent advances in Neural ODEs on constrained manifolds. We reformulate the task of learning causal structures at each time step as solving the solution trajectory of a Neural ODE on the directed acyclic graph (DAG) manifold. To accommodate high-dimensional causal structures, we extend DyCAST by learning the temporal dynamics of the hidden state for contemporaneous causal structure. Experiments on both synthetic and real-world datasets demonstrate that DyCAST achieves superior or comparable performance compared to existing causal discovery models.

Weakly-supervised fine-grained human parsing, which decomposes the human body into several parts and various fashion items only with some easier labels, poses a more challenging visual task and cannot be well solved by general weakly-supervised approaches. In this case, we first explore the feasibility of utilizing point-level labels to address this task. Toward this, we propose the prior-structure driven weakly-supervised learning for fine-grained human parsing. Following previous practices, we design a pseudo label initialization mechanism to produce high-quality pixel-level pseudo labels by utilizing the powerful image segmentation model Segment Anything Model (SAM). Then we propose the Feature Propagation based on Prior-Structure (FPPS) module which formalizes prior-structure knowledge as an adjacency matrix constructed from superpixel and emploies a learnable Graph Neural Network (GNN) as the feature propagator. FPPS can optimize the features of unlabeled pixels to enhance the weakly-supervised learning. The framework further designs the Refinement Pseudo Label (RPL) strategy to generate denser supervision from past sub-optimal models. To the best knowledge, this work is the first attempt to perform fine-grained human parsing in a weakly-supervised manner. We conduct extensive experiments on two challenging fine-grained datasets, including ATR and LIP. Experimental results show that the proposed weakly-supervised method yields a comparable result to strongly-supervised methods and even outperforms other state-of-the-art approaches in semi-supervised human parsing tasks.

Mobile Crowd Sensing (MCS) represents an emerging paradigm for collecting sensory data, leveraging the extensive sensing capabilities of widely used mobile devices to execute sensing tasks. Among the array of challenges facing current MCS systems, the incentive mechanism for data requesters and participants consistently stands out as a paramount concern. Existing incentive mechanisms often rely on model-based approaches, assuming a certain degree of prior knowledge about the MCS system, such as expected pricing for data requesters and participants. However, these assumptions are impractical in real-world scenarios. To address this challenge, we endeavor to explore a wholly model-free incentive mechanism. Specifically, we propose a Simulated Annealing Deep Q-learning (SADQ-learning) algorithm to dynamically generate the pricing policy for the sensing platform. Furthermore, to accommodate diverse incentive needs, we devise three distinct incentive modes: one focuses on maximizing the profit of the sensing platform, another dedicates to maximizing the successful matching amount of sensing tasks, and an equilibrium mode seeks a balance between the aforementioned objectives. Finally, numerical results demonstrate the superiority of SADQ-learning through comparisons with baseline algorithms.

Assembly sequence planning (ASP) is the most important process in the product lifecycle. This paper aims to propose a parallel assembly sequence planning method based on sparrow search algorithm (SSA) to improve the assembly quality and efficiency. The proposed method divides the problem into assembly unit division (AUD) part and ASP part. In AUD part, we use the markov clustering algorithm (MCA) to solve the AUD problem. In ASP part, the ASP evaluation system is designed for the assembly part sequence planning (APSP) problem and assembly unit sequence planning (AUSP) problem, and the discrete artificial sparrow search algorithm (DASSA) is proposed for solving the discrete APSP problem. The results of comparison experiment verify the effectiveness of the proposed method. Compared with genetic algorithm (GA) and particle swarm optimization (PSO), the proposed DASSA gets the optimal results and fastest convergence. The method proposed solves ASP problem to efficiently perform parallel assembly sequence planning and reduce production time, making it cost-effective in intelligent manufacturing.

Optical Communication Simulation Task Allocation (OCSTA) constitutes an interdisciplinary quandary. This paper proposes a swarm intelligence approach harnessing a Discrete Dung Beetle Optimizer (DDBO) algorithm with a globally equilibrated strategy to offer a resolution avenue for this intricate conundrum. Firstly, we provide a comprehensive exposition of the mathematical model underpinning the OCSTA, which employs the spatial positioning of the population to articulate diverse allocation solutions. Then, a weighted random selection technique is used to generate initial solutions. Thirdly, a predator avoidance strategy is introduced to facilitate updates in the dung beetle position. Finally, the global equilibrium mechanism with the swap operator is exploited to harmonize the exploratory and exploitative capabilities, thereby further augmenting the quality of the solutions. We conducted several simulation experiments1across various distinct task load scenarios, and statistical tests are employed to evaluate the significant differences between the proposed algorithm and other state-of-the-art methods. The outcomes revealed that the DDBO solution yielded an improvement of approximately 15.8%, which underscores the competitiveness and robustness of solving the OCSTA.

Directed Acyclic Graphs (DAGs) are widely utilized across various domains for tasks such as graph algorithms, data flow analysis, program optimization, and machine learning. Representing DAGs using General-purpose Programming Languages (GPLs) or Data Serialization Formats (DSFs) can lead to complex and obscure expressions, making it challenging to comprehend and manage the codebase. Domain-Specific Languages (DSLs) offer a more tailored approach, but come with limitations and development overhead. This paper introduces the Open Customizable Modeling Language (OpenCML), a universal DAG modeling language specification that aims to provide a standardized and customizable framework for modeling and scripting DAGs. Evaluations demonstrate that OpenCML offers expressive power, customizability, and interoperability, simplifying the learning process and providing a powerful solution for DAG modeling and scripting.

Despite the astounding progress made in semi-supervised learning (SSL) and imbalanced supervised learning (ISL), there has been little attention devoted to the research of imbalanced semi-supervised learning (ISSL). The ‘Matthew effect’, a phenomenon where a disparity in data representation becomes more severe in a class-imbalanced dataset during training, could be amplified in a semi-supervised setting. In this study, we addressed two key challenges in ISSL: maintaining the reliability of pseudo-labels and ensuring a balanced representation of features. Specifically, we propose a class-aware feature-diffusion constraint and reliable pseudo-labeling (DCRP) framework to address these issues. In the DCRP, we counteract the overconfidence problem of softmax by adding an extra class to the typical K class problem without the need for additional parameters. Moreover, we introduced a flexible class-aware feature diffusion constraint in the feature extractor, promoting a more balanced feature diversity. Experimental validations on various datasets, such as CIFAR10-LT, CIFAR100-LT, SVHN-LT, and Small ImageNet-127, demonstrated consistent improvements in accuracy with our DCRP method. In particular, we achieved a steady improvement in accuracy of approximately 1% under the newly published ACR prototype across most settings. The code is available athttps://github.com/guoxiaoyuatbjtu/DCRP.

Background:Multi-source cross-project defect prediction (MSCPDP) attempts to transfer defect knowledge learned from multiple source projects to the target project. MSCPDP has drawn increasing attention from academic and industry communities owing to its advantages compared with single-source cross-project defect prediction (SSCPDP). However, two main problems, which are how to effectively extract the transferable knowledge from each source dataset and how to measure the amount of knowledge transferred from each source dataset to the target dataset, seriously restrict the performance of existing MSCPDP models.Objective:In this paper, we propose a novel multi-source transfer weighted ensemble learning (MASTER) method for MSCPDP.Method:MASTER measures the weight of each source dataset based on feature importance and distribution difference and then extracts the transferable knowledge based on the proposed feature-weighted transfer learning algorithm. Experiments are performed on 30 software projects. We compare MASTER with the latest state-of-the-art MSCPDP methods with statistical test in terms of famous effort-unaware measures (i.e., PD, PF, AUC, and MCC) and two widely used effort-aware measures (Popt20% and IFA).Result:The experiment results show that: 1) MASTER can substantially improve the prediction performance compared with the baselines, e.g., an improvement of at least 49.1% in MCC, 48.1% in IFA; 2) MASTER significantly outperforms each baseline on most datasets in terms of AUC, MCC,Popt20% and IFA; 3) MSCPDP model significantly performs better than the mean case of SSCPDP model on most datasets and even outperforms the best case of SSCPDP on some datasets.Conclusion:It can be concluded that 1) it is very necessary to conduct MSCPDP, and 2) the proposed MASTER is a more promising alternative for MSCPDP.

Age of information (AoI) is an important metric used to quantify the freshness of data. By utilizing resources of the edge server near the source nodes, mobile edge computing (MEC) can speed up the processing of information updates and ensure data freshness. However, the resources of the edge server are usually limited, so it is necessary to study the resource allocation strategy to ensure data freshness and optimize the profit of the edge server. In this paper, we propose a game-theoretic approach for resource allocation in mobile edge computing to guarantee information freshness. First, with the purpose of ensuring information freshness, we formalize the problem as minimizing the computational cost of source nodes and maximizing the profit of the edge server. Then, we introduce a two-stage dynamic game model to simulate the competitive process. We further transform the resource allocation problem into a knapsack problem and propose an iterative resource allocation algorithm based on dynamic programming. Experimental results show that the proposed algorithm can obtain a Nash equilibrium and maximize the profit of the edge server while ensuring information freshness.

The recently proposed class-imbalanced semi-supervised learning (CISSL) algorithms achieved impressive performance by effectively leveraging unlabeled data. However, these algorithms often rely on a pre-defined fixed confidence threshold to filter unlabeled data during training, which overlooks the varying learning dynamics across different classes in class-imbalanced scenarios. Consequently, valuable data could be discarded, leading to degraded performance on minority classes. To tackle this issue, we introduce a novel method called Class-Adaptive Threshold (CAT), which dynamically defines and adjusts the confidence threshold based on the learning status of each class. The core idea of CAT is to iteratively update the thresholds for different classes at each time step, enabling us to fully exploit valuable information that would otherwise be ignored using fixed threshold algorithms. Importantly, CAT does not introduce any additional inference processes. In our experiments, the proposed algorithm achieves state-of-the-art performance on various class-imbalanced datasets. Furthermore, we show that CAT can be seamlessly integrated into the renowned CISSL algorithm, resulting in a remarkable boost in their performance.

Forecasting crowd movement accurately across an urban area is crucial for efficient traffic control and ensuring public security. Current methods involve transforming the city’s roadmap into a grid-based map, enabling Convolutional Neural Networks (CNNs) or Graph Convolutional Networks (GCNs) to capture spatio-temporal relationships efficiently. However, this approach overlooks the connection between irregularly shaped real-world areas, which can be categorized into various functional zones. In this article, we introduce a novel approach for predicting urban crowd flow named STHGN, which utilizes hypergraph convolutional networks. By constructing 3-level hypergraphs from irregular areas and adopting Hyper-GCN, we capture mobility among irregular regions. We construct the hypergraphs based on hour, day, and week, simultaneously using gated-based mechanisms to fuse various embeddings. We evaluate the efficacy of our model by contrasting it with 11 other approaches, including the most sophisticated STGs. After conducting numerous experiments, we find that STHGN outperforms these methods with higher accuracy, resulting in a reduction of approximately 6-9% in mean absolute error (MAE) for crowd flow prediction.

Surface defects of railway trackside equipment pose a serious risk on the safety of railway transportation systems. Image-based surface defect detection methods have made significant progress. However, the image background of trackside equipment is complex, and there is a large amount of noise, which makes existing methods inadequate in accurately detecting small surface defect regions. To tackle with this issue, we propose a coarse-to-fine parallelizable surface defect detection approach to hierarchically detect the defects of trackside equipment. Firstly, a detection network is designed to locate and extract trackside equipment, which aims at roughly focusing the detection field from the original image to the region of interest of individual trackside equipment. Then, a novel semantic segmentation network is proposed to segment the major components of trackside equipment, so as to further finely focus on the defect regions. We apply multiple segmentation networks to parallelly segment various trackside equipment. In the segmentation network, a dense feature enhancement method is introduced to strengthen the high-level semantic information, and a feature partitioning enhancement strategy is designed to improve the segmentation performance for small defect regions. Finally, according to the visual characteristics of the segmentation output, we propose a defect recognizer to discriminate the defects. Extensive experimental results demonstrate that the proposed surface defect detection approach achieves higher accuracy for trackside equipment.

Click-Through Rate(CTR) prediction aims to predict the possibility of users clicking on products, which has become the core task of advertising recommendation systems. Due to the richness of user historical behavior, a key to making effective prediction is to capture users' diverse interests from historical behavior. An efficient way to do this is to perform dot product of behavior and target embedding with attentive neural networks. To better model the users' diverse interests, our proposed disentangled interest extraction block decouples the unary terms modeling the impact of user behavior sequence from pairwise interactions. Specifically, the decoupled pairwise term can learn the pure pairwise interactions, whereas the unary term models the impact of behavior sequence on each target items. Meanwhile, our model emphasizes both high- and low-order feature interactions by combining Attentional Factorization Machines(AFMs) with deep learning. This work intends to accomplish our goal by proposing a novel architecture Deep Disentangled Interest Network(DDIN). We conduct comprehensive experiments on Movielens dataset and Amazon electronic dataset. The results demonstrate the effectiveness of DDIN which is superior to some state-of-art models by up to 26.271 %.

Long-term time-series forecasting is an extensive research topic and is of great significance in many fields. However, the task of long-term time-series forecasting is accompanied by the problem of increasing cumulative error and decreasing time correlation. To overcome these shortcomings, this article proposes a prediction framework based on the nonlinear fuzzy information granule (NFIG) series, which can boost the long-term performance of most predictors. First, we propose the representation of the NFIG for the first time, replacing the linear core lines with nonlinear time-dependent curves. Second, we propose a temporal window splitting algorithm based on curvature equations and weighted directed graphs, which can not only merge temporal windows with the same trend but also cointegrate incremental data. Finally, the nonlinear trend fuzzy granulation can be employed as a data preprocessing module for various time-series predictors to achieve a better long-term forecasting performance. As a typical time-series forecasting task, the precise long-term forecast of traffic flow data can relieve the overburdened traffic system and improve the traffic environment to a certain extent. Thus, the proposed method is employed for the long-term traffic flow forecasting. Compared with existing forecasting models, which achieves superior performances.

Deep learning technologies have been applied in various computer vision tasks in recent years. However, deep models suffer performance decay when some unforeseen data are contained in the testing dataset. Although data enhancement techniques can alleviate this dilemma, the diversity of real data is too tremendous to simulate. To tackle this challenge, we study a scheme for improving the robustness and efficiency of the deep network training process in visual tasks. Specifically, first, we build positive and negative sample pairs based on a class-sensitive strategy. Then, we construct a feature-consistent learning strategy based on contrastive learning to constrain the representations of interclass features while paying attention to the intraclass features. To extend the effect of the consistent strategy, we propose a novel contrastive Jensen-Shannon divergence consistency loss (JS loss) to restrict the probability distributions of different sample pairs. The proposed scheme successfully enhances the robustness and accuracy of the utilized model. We validated our approach by conducting extensive experiments in the domains of model robustness and few-shot object detection (FSOD). The results showed that the proposed method achieved remarkable gains over state-of-the-art (SOTA) methods. We obtained a 3.2% average improvement over the best-performing FSOD method.

Context: Software defect number prediction (SDNP) helps allocate limited testing resources by ranking software modules according to the predicted defect numbers. However, the highly skewed distribution of defects greatly degrades the performance of SDNP models by preventing SDNP models from ranking software modules accurately. Objective: This paper introduces a novel subspace hybrid sampling ensemble (SHSE) method based on feature subspace construction, hybrid sampling , and ensemble learning for building high-performance SDNP models. Method: Specifically, we first construct a series of feature subspace to ensure the diversity of base learners. In each of feature subspace, we then use the proposed hybrid sampling method to balance the training subset without losing too much information and introducing lots of noisy data caused by only using undersampling or oversampling techniques. Finally, we train each base learner and combine them by using the proposed weighted ensemble strategy. Experiments are performed on 27 public defect datasets. We compare SHSE with five state-of-the-art resampling-based models and four zero-inflated/hurdle models in terms of the ranking performance measure fault-percentile-average (FPA). To demonstrate the effectiveness of SHSE, two statistical testing methods including Wilcoxon Signed-rank test and Scott–Knott Effect Size Difference test are utilized. Cliff’s δ is also computed for quantifying the difference when there is significant difference between SHSE and each baseline. Results: The experimental results show that SHSE significantly outperforms the baselines and improves the performance over each baseline with as least medium effect size on most datasets. On average, SHSE improves the performance over the resampling-based methods by 8.7% ∼ 14.4% and the zero-inflate/hurdle models by 10.3% ∼ 15.2%. Conclusion: It can be concluded that SHSE is a more promising alternative for software defect number prediction.

Visual-based unsupervised learning[1]–[3]has emerged as a promising approach in estimating monocular depth and ego-motion, avoiding intensive efforts on collecting and labeling the ground truth. However, they are still restrained by the brightness constancy assumption among video sequences, especially susceptible with frequent illumination variations or nearby textureless surroundings in indoor environments. In this article, we selectively combine the complementary strength of visual and inertial measurements, i.e., videos extract static and distinct features while inertial readings depict scale-consistent and environment-agnostic movements, and propose a novel unsupervised learning framework to predict both monocular depth and ego-motion trajectory simultaneously. This challenging task is solved by learning both forward and backward inertial sequences to eliminate inevitable noises, and reweighting visual and inertial features via gated neural networks in various environments or with user-specific moving dynamics. In addition, we also employ structure cues to produce scene depths from a single image and explore structure consistency constraints to calibrate the depth estimates in indoor buildings. Experiments on the outdoor KITTI data set and our dedicated indoor prototype reveal that our approach consistently outperforms the state of the art on both depth and ego-motion estimates. To the best of our knowledge, this is the first work to fuse visual and inertial data without any supervision signals for monocular depth and ego-motion estimation, and our solution remain effective and robust even in textureless indoor scenarios.

Semi-supervised learning (SSL) has been studied widely in standard benchmark datasets; however, real-world data often exhibit class-imbalanced distributions, which pose significant challenges for deep semi-supervised models. To address this issue, we design a three-stage learning framework, 3LPR, by combining unsupervised feature extraction, graph-based Label Propagation, and mixed data augmentation (MDA)-based label Reassignment. Specifically, we first explore the performance of supervised and unsupervised learning for feature extraction of class-imbalanced data and then establish our first stage of feature extraction through unsupervised learning. Then, we adopt graph network-based offline label propagation and sieving to effectively expand the labeled set to overcome the excessive label bias in the classifier during the training process. Finally, we propose a label reassignment (LRA) algorithm for class-imbalanced semi-supervised learning (CISSL) to train the expanded dataset, where the MDA strategy is adopted but with the label reassigned. The experimental results demonstrate that the proposed 3LPR framework for CISSL outperforms other state-of-the-art methods on various datasets.

With the rapid development of intelligent transportation, automated traffic surveillance is considered as an important component. In the field of traffic surveillance, it is particularly important to achieve robust and real-time tracking of vehicles in complex scenes. In this paper, a robust real-time vehicle tracking model namedNoisyOTNetis proposed, which formulates tracking as reinforcement learning with parameter space noise. In this formulation, the exploration ability of the model is enhanced to improve the robustness of tracking. Specifically, we develop a new implementation for noisy network based on deep deterministic policy gradients (DDPGs) with parameter noise, which can better cope with the tracking task and directly predict the tracking result. To improve the tracking accuracy in complex conditions, e.g. fast motion and large deformation, this paper presents an adaptive update strategy that can exploit the vehicle spatial-temporal information based on Upper Confidence Bound (UCB) algorithm by exploiting. Moreover, as for the recovery of the lost target, a relocation algorithm based on incremental learning is developed. The results of extensive experiments demonstrate that the proposed NoisyOTNet can effectively track vehicles in complex scenes and achieve competitive performance compared to the state-of-the-art methods.

Estimating the origin-destination travel time is a fundamental problem in many location-based services for vehicles, e.g., ride-hailing, vehicle dispatching, and route planning. Recent work has made significant progress to accuracy, but they largely rely on GPS trajectories which are too coarse to model many personalized driving behaviors, e.g., differentiating novice and veteran drivers. In this paper, we propose Customized Travel Time Estimation (CTTE) that fuses GPS trajectories, smartphone inertial data, and road network within a deep recurrent neural network. It constructs a road link traffic database with topology representation, speed statistics, and query distribution. It also calibrates inertial readings, estimates the arbitrary phone’s pose in car, and detects multiple aggressive driving events (e.g., bump judders, sharp turns, sharp slopes, frequent lane shifts, overspeeds, and sudden brakes). Finally, we demonstrate our solution on two typical transportation problems, i.e., predicting traffic speed at holistic level and estimating customized travel time at personal level, within a multi-task learning structure. Experiments on two large-scale real-world traffic datasets from DiDi platform show our effectiveness compared with the state-of-the-art.

With the development of edge computing (EC), more and more tasks are offloaded to edge servers (ESs). However, faced with a huge number of users offloading tasks to ESs, how to allocate resources reasonably and reduce the response time of the system are problems worth studying. In this paper, we design an online auction algorithm to deal with those two issues at the same time. We first introduce four task value functions to model the sensitivity to the delay of different tasks. Then, we construct a three-layer EC model. Based on it, we define the resources allocation problem as a social welfare (SW) maximization problem, which is NP-hard. To solve this problem, we utilize the master-dual technique to transform it into an online auction problem. Finally, an algorithm considering task classification is proposed, which realizes both resource allocation and latency reduction in a polynomial time. Experiment results show that our approach reduces the scheduling latency by an average of 38% while maintains SW at the same time.

Human pose estimation plays an important role in human action recognition, human-computer interaction, animation. Most existing methods commonly utilize cascaded pyramid or stacked hourglass network to fuse multi-scale feature from different levels, which greatly enhances the performance but brings a tremendous amount of computing, so it is difficult to achieve real-time human pose estimation. In this paper, we propose a newly-designed network named Complementary Feature Pyramid Network (CFPNet) for human pose estimation with a focus on efficient multi-scale feature generate and fusion method. CFPNet extends the range of receptive fields for each network layer with the help of Feature Mix Bottleneck (FMB) block which constructs hierarchical connections and mixes multiple receptive fields features in a single bottleneck block. In order to reduce the redundant gradient information during the network optimization and construct a lightweight network, Cross Stage Partial (CSP) connection is introduced into the CFPNet. Complementary Feature Fusion (CFF) block is proposed, which can adaptively select complementary information from different levels for fusion to maximize the effective feature in the output of CFPNet. Through the above improvements, CFPNet comprises more affluent multi-scale feature and lower model complexity. Especially, CFPNet-101 achieves the 72.3% AP at 31.7 FPS on the MS COCO dataset only with 1.96 GFLOPs and 10.5M Params. Compared with the existing methods, CFPNet has competitive accuracy and can run in real-time.

Although vehicle location-based services are prevalent outdoors, we are back into darkness in many GPS blocked environments, such as tunnels, indoor parking garages, and multilevel flyovers. Existing smartphone-based solutions usually adopt inertial dead reckoning to infer the trajectory, but low-quality inertial sensors in phones are plagued by heavy noises, causing unbounded localization errors through double integrations for movements. In this article, we propose VeTorch, a smartphone inertial odometry that devises an inertial sequence learning framework to track vehicles in real time when GPS signal is not available. Specifically, we transform the inertial dynamics from the phone to the vehicle regardless of the arbitrary phone's placement in the car and explore a temporal convolutional network to learn the vehicle's moving dependencies directly from the inertial data. To tackle the heterogeneous smartphone properties and driving habits, we propose a federated learning-based active model training mechanism to produce customized models for individual smartphones, without incurring user privacy issues. We implement a highly efficient prototype and conduct extensive experiments on two large-scale real-world traffic data sets collected by a modern ride-hailing platform. Our results outperform the state-of-the-art vehicular inertial dead-reckoning solutions on both accuracy and efficiency.

This article deals with H∞state estimation of neural networks with mixed delays. In order to make full use of delay information, novel delay-product Lyapunov-Krasovskii functional (LKF) by using parameterized delay interval is first constructed. Then, generalized free-weighting-matrix integral inequality is used to estimate the derivative of LKF to reduce the conservatism. Also, a more general activation function is further applied by combining with parameterized delay interval in order to obtain a more accurate estimator model. Finally, sufficient conditions are derived to confirm that the estimation error system is asymptotically stable with a prescribed H∞performance. Numerical examples are simulated to show the benefits of our proposed method.

Caching has become more and more important in mobile video delivery with the development of communication technology, which can reduce redundant data transmissions and decrease costs. In this paper, we consider a multiple bitrate caching system consisting of a network service provider (NSP) in charge of edge cloud with caching and processing capacity, a set of video providers (VPs), and multiple mobile users. The NSP leases the storage space of the edge cloud to VPs and VPs place their videos in edge cloud to provide better services for mobile users. The competition between NSP and VPs is modeled as a Stackelberg game. We first solve the knapsack problem to find the optimal cache placement strategy for each VP based on the allocated space to maximize its profit and then adjust the allocation strategy to maximize the profit of NSP. Moreover, we develop a dynamic programming algorithm and a multiple bitrate caching algorithm to find the Stackelberg equilibrium (SE). The numerical results show the effectiveness of proposed algorithms on pricing and cache placement.

With the development of deep learning and the emergence of massive video data, object tracking has great application prospects in many fields. However, most tracking algorithms can hardly get top performance with real-time speed. In this paper, we improved tracking model based on adversarial learning and to accelerate feature extraction we proposed an efficient and accurate method. We also present a Precise ROI Pooling (PrROIPooling) based algorithm for extracting more accurate representations of targets. Furthermore, a novel regularization term is defined to ensure the similarity between the generated features and the real features. Finally, the improved objective function with modulating factors is designed to handle the problem of imbalance in the number of positive and negative samples. Extensive experiments on three datasets have demonstrated our effectiveness and achieved competitive results compared with state-of-the-art methods.

Abstract Due to the inherent nonlinear nature of data, traditional linear methods have some limitations in finding the intrinsic dimensions of motion capture (Mo‐cap) data. Mo‐cap data are more in line with the characteristics of the manifold. Assuming that the data are initially a low‐dimensional manifold and uniformly sampled in high‐dimensional Euclidean space, manifold learning recovers low‐dimensional manifold structures from high‐dimensional sampled data. This paper proposes an automatic segmentation method based on geodesics by introducing a Riemannian manifold. We convert Mo‐cap data from Euler angles into quaternions, calculate the intrinsic mean of the motion sequence, hemispherize quaternions, and use logarithmic and exponential mapping to calculate geodesic distances instead of quaternions. The experimental results show that the algorithms can achieve automatic segmentation and have a better segmentation effect.

In this paper, we propose a novel tracking method by formulating tracking as a correlation filtering as well as a ridge regression problem. First, we develop a tight correlation filter-based tracking framework from the signal detection perspective. In this formulation, the correlation filter is set as the same size as the target, which can make full use of the relations of the adjacent image patches and effectively exclude the influence of the background. Specifically, we point out that the novel correlation filter model can be regarded as the ridge regression model which takes into account the different importance of the samples and has the consistent objective with tracking. Second, we focus on the scale variation problem in tracking. By making use of the spatial structure of the correlation filter, the multiscale filter banks can be generated via interpolation to handle the scale estimation problem easily. Third, we present a novel distance importance-based confidence calculation model to determine the final tracking result, which not only makes use of the fine discriminability of the correlation filter but also takes the distance importance of the candidate samples into account to alleviate the impact of similar distractors. Experimental results demonstrate that our method is superior to several state-of-the-art trackers and many other correlation filter-based methods in the benchmark datasets.

In traditional exemplar-based image inpainting algorithm, the confidence value will rapidly decrease to zero as the inpainting process progresses.As a consequence, it will lead to unreliable result of the priority calculation and wrong direction of the process.In addition, traditional methods usually use the sum of squared differences (SSD) criterion to search the optimal matching block.Since the matching criterion is single and the precision is limited, the process is easy to produce mismatch.In order to solve the above problems, an improved algorithm has been proposed in this paper.First, we proposed a new confidence update algorithm through replacing the previous linear function form by using a logarithmic function form, which can suppress the phenomenon that the confidence attenuation is too fast and improve the accuracy of guiding and inpainting direction.Then, we combine the physical distance between blocks and traditional SSD matching criterion to improve matching accuracy.The experimental results show that the algorithm overcomes the shortcomings of the traditional algorithm and provides higher quality image restoration effects and better visual effects.

Since Correlation Filter appeared in the field of video object tracking, it is great popular due to its excellent performance.The Correlation Filter based tracking algorithms are very competitive in terms of accuracy and speed as well as robustness.However, there are still some fields for improvement in the Correlation Filter based tracking algorithms.First, during the training of the classifier, the background information that can be utilized is very limited.Moreover, the introduction of the cosine window further reduces the background information.These reasons reduce the discriminating power of the classifier.This paper introduces more global background information on the basis of the DCF tracker to improve the discriminating ability of the classifier.Then, in some complex scenes, tracking loss is easy to occur.At this point, the tracker will be treated the background information as the object.To solve this problem, this paper proposes a novel re-detection component.Finally, the current Correlation Filter based tracking algorithms use the linear interpolation model update method, which cannot adapt to the object changes in time.This paper proposes an adaptive model update strategy to improve the robustness of the tracker.

Since Correlation Filter appeared in the field of video object tracking, it is very popular due to its excellent performance. The Correlation Filter-based tracking algorithms are very competitive in terms of accuracy and speed as well as robustness. However, there are still some fields for improvement in the Correlation Filter-based tracking algorithms. First, during the training of the classifier, the background information that can be utilized is very limited. Moreover, the introduction of the cosine window further reduces the background information. These reasons reduce the discriminating power of the classifier. This paper introduces more global background information on the basis of the DCF tracker to improve the discriminating ability of the classifier. Then, in some complex scenes, tracking loss is easy to occur. At this point, the tracker will be treated the background information as the object. To solve this problem, this paper introduces a novel re-detection component. Finally, the current Correlation Filter-based tracking algorithms use the linear interpolation model update method, which cannot adapt to the object changes in time. This paper proposes an adaptive model update strategy to improve the robustness of the tracker. The experimental results on multiple datasets can show that the tracking algorithm proposed in this paper is an excellent algorithm.

Recently, various regression based tracking methods have achieved great success. However, in most of these methods, all of the extracted features are made use of to represent the object without feature selection. In this paper, we propose a novel tracking method based on elastic net regression with adaptive weights. On one hand, tracking is formulated as an elastic net regression problem which can not only make full use of the spatial information, but also automatically select features to alleviate the influence of the unstable or inaccurate points. On the other hand, the weights of the ℓ1-norm and ℓ2-norm regularization in the regression model are adaptively adjusted to better improve the performance. Experimental results in the benchmark dataset demonstrate that the proposed adaptive elastic net regression based tracking method can achieve desirable tracking performance.

Resilience is the property that enables a system to continue operating properly when one or more faults occur. Nowadays, as software systems become more and more complex, their hardware execution platforms also become more heterogenous with larger scale. Software systems may fail due to some faults such as node breakdown, communication failure, or data processing failure. In this paper, we propose a ring-based resilience mechanism, which implements fault detection and recovery. (1) To solve the problem that the central server may have high burden of network traffic, we design a ring-based heartbeat algorithm for crash fault detection. (2) We also design a light-weight recovery mechanism to recover from crash faults as compared with the current system-specific mechanisms. To evaluate our mechanism, we use a 3D-online game system as a case study. By injecting faults, we test the effectiveness and overhead of the proposed mechanism. Compared with other mechanisms, the experimental results show that our mechanism can support resilience very well and is better at dealing with the crash fault caused by high cluster workload with acceptable overhead.

Improving image resolution has broad applications and is an important research topic. Recently, a hybrid method Adaptive Sparse Domain Selection (ASDS) combining a reconstruction‐based method and an example‐based method has been proposed to take advantage of the two, but may not reconstruct sufficient details. In this study, the authors propose to improve ASDS: Zeyde's method is first used to obtain an intermediate image with high‐frequency details, and then the obtained image is used to replace the autoregressive model of ASDS as the example‐based term. In addition, the authors may split the input image into patches and use different parameter settings for the patches of different amount of details. Experimental results demonstrate the improved hybrid methods can produce high‐quality images quantitatively and perceptually.

Abstract In order to conveniently classify, retrieve, and synthesize human motion, motion capture (MoCap) data need to be properly segmented into distinct behaviors. In this paper, we propose a novel automated segmentation method based onposture histograms in sliding window. Firstly, a set of new posture features are proposed and defined to construct theposture histogram, which is a new compact representation of behavioral features. Then, by executing thesliding window, especially in this paper, the behavior features are analyzed in subsequence level to reduce noise sensitivity. We open up a novel way to tune sliding window by studying steady states of human behaviors, so that conspicuous and stable behavioral features can be obtained. Finally, by analyzing the clustering property of posture histograms of the subsequences, the behavior segmentation problem is tactfully simplified to the detection ofoutlier subsequence. In particular, thelocal outlier factoralgorithm is adopted to solve outlier subsequence detection, and good results are achieved. Extensive experiments are conducted on 14 pieces of multi‐bcase, and the experimental results demonstrate that our proposed method outperforms other state‐of‐the‐art ones. Copyright © 2016 John Wiley & Sons, Ltd.

Cloud computing has been a mainstream solution for the processing and storage of mass data, as well as an exciting area for research. As a novel business model, cloud computing has dramatically changed the provision of services and IT capacity by means of the advanced techniques. In recent years, with the increasing research interests and rapid growth of publications, some review papers provide detailed analysis on cloud computing area. In this paper, a bibliometric-based approach is presented and implemented to quantitatively review the progress in global cloud computing research with the related literature during 2007–2013 from the databases of Science Citation Index Expanded (SCI-E), Conference Proceedings Citation Index–Science (CPCI-S), and IEEEXplore. Our work is motivated by the purpose of tracing global advancement in terms of research content, geographic distribution and issue time of the related publications, rather than a specific technological area in cloud computing research. By investigating the characteristics of publications such as keywords, output, geographic distribution and affiliation, we draw some valuable conclusions to guide the further research. The experimental results show that the top 5 active research points of cloud computing concentrate on virtualization, security, mobile cloud, distributed computing, and scheduling. From the location-time aspect, China, USA, and India have published most of the papers, dominate cloud computing research and keep a high level on the international research cooperation. And there is a great increase in publication outputs especially in China and USA. Meanwhile, the analysis results demonstrate the top 3 high-cited research institutes of the University of Melbourne, University of California. Berkeley and University of Vienna in cloud computing research. The mobile cloud will be a future research hotspot and promising application field.

Crowd simulation explores crowd behavior in virtual environments, which has been extensively studied in many areas, such as safety and civil engineering, transportation, social science, and entertainment industry. In this paper, an improved potential field method is proposed to achieve the real-time crowd simulation, which is composed of the global navigation with Dijkstra's algorithm and the potential field based local navigation. First, a region separation is performed to divide the environment into a set of triangles, and thus a topological graph can be built with the triangles as vertices. Then a velocity-density model is introduced for improving the speed controlling mechanism and solving the "maximum speed dilemma" which means the velocity of an individual derived by potential field will be stuck into the maximum due to the ill speed control. Since the movement of an individual in the crowd is influenced by the socio-psychological forces, the individuals' actions express the group attributes. In order to represent the group attributes in the crowd, the repulsive potential function is improved in this paper. Experiments have been carried out and the results show that the improved potential field based method can simulate the crowd in real time and avoid the "maximum speed dilemma".

Creating realistic virtual humans has been a challenging objective in the areas of computer science research and technology industry. While there are a number of aspects to create realistic virtual humans, this paper focuses on the comprehensive integrated framework of modeling virtual humans with high level autonomy, which aim to reproduce human-like believable behaviors and movements of virtual humans in virtual environment. In the framework, the perception module enables virtual human to explore the virtual environment and gets vision and audition information; the decision networks based behavioral decision-making module allows virtual human react appropriately to the perceived surrounding environment; the hierarchical movement animation control module is designed to generate autonomous character navigation and realistic motions for character animation in virtual environment. The integrated framework presented is tested in the simulated virtual environment.

This paper presents a 3D object classification system based on volumetric parts. As the constituents of 3D object, the parts are described by superquadric-based geons, which enables a more compact 3D object representation. In the developed classification system, the improved interpretation tree method is implemented for classification, where a set of novel integrated features and corresponding constraints are proposed, which not only reflect individual parts’ shape, but model’s topological structure among 3D parts. The constraints are used to define efficient interpretation tree search rules, and the feasible correspondences of unknown object data and the stored models are obtained. Then a similarity measure computation algorithm is proposed to evaluate the shape similarity of the correspondence. The classification system can achieve both whole match and partial match between unknown object data and 3D models with shape similarity ranks; particularly, focus match can be accomplished, in which different key parts may be labeled and all the matched models with corresponding key parts can be obtained. The performance of the presented 3D object classification system is evaluated with a series of experiments.