In cloud computing, static resource configurations create a trade-off: tenants overprovision to avoid resource starvation, causing inefficiency and cost, while providers suffer low utilization despite high allocations. To improve efficiency, providers often use overcommitted environments where multiple workloads share hosts, but this leads to interference and potential Quality-of-Service (QoS) violations. This paper introduces a realtime dynamic control framework that mitigates interference by adaptively pinning workloads to CPU groups. Using deep reinforcement learning (DRL) with the Proximal Policy Optimization (PPO) algorithm, an intelligent agent continuously adjusts CPU pinning based on real-time feedback to maintain Service- Level-Agreement (SLA) compliance. Experiments under two optimization objectives—overall-performance-first and priorityperformance- first—show that the proposed approach improves overall QoS by$\approx$25% compared with static pinning. When prioritization is enabled, high-priority workloads gain significant performance improvements while lower-priority ones remain within SLA limits. These results demonstrate that a DRL-based CPU-pinning strategy effectively manages resource contention in overcommitted clouds, enhancing utilization while upholding tenant SLAs.

Heterogeneous domain adaptation (HDA) is expected to achieve effective knowledge transfer from a label-rich source domain to a heterogeneous target domain with scarce labeled data. Most prior HDA methods strive to align the cross-domain feature distributions by learning domain invariant representations without considering the intrinsic semantic correlations among categories, which inevitably results in the suboptimal adaptation performance across domains. Therefore, to address this issue, we propose a novel semantic correlation transfer (SCT) method for HDA, which not only matches the marginal and conditional distributions between domains to mitigate the large domain discrepancy, but also transfers the category correlation knowledge underlying the source domain to target by maximizing the pairwise class similarity across source and target. Technically, the domainwise and classwise centroids (prototypes) are first computed and aligned according to the feature embeddings. Then, based on the derived classwise prototypes, we leverage the cosine similarity of each two classes in both domains to transfer the supervised source semantic correlation knowledge among different categories to target effectively. As a result, the feature transferability and category discriminability can be simultaneously improved during the adaptation process. Comprehensive experiments and ablation studies on standard HDA tasks, such as text-to-image, image-to-image, and text-to-text, have demonstrated the superiority of our proposed SCT against several state-of-the-art HDA methods.

Since the advent of cloud computing, Elastic Computing (EC) has become the standard architecture for resource allocation and scheduling. EC typically allocates computing resources based on predefined specifications, such as virtual machine or container flavors. However, these flavors are often constrained by fixed CPU-to-memory ratios, which frequently fail to match the actual resource needs of applications. As a result, cloud providers experience high resource allocation rates nearing saturation ($> $80%) but with low utilization ($< $25%). This study introduces Flexible Computing (FC), a novel approach to resource allocation and scheduling. Unlike EC, FC allocates resources based on an application resource usage profile, derived from the historical resource consumption of workloads, rather than relying on fixed specifications. Additionally, FC incorporates a real-time performance degradation detection mechanism to address performance issues caused by the noisy-neighbor effect when colocated workloads interfere with each other. FC dynamically adjusts resource allocation according to actual usage, ensuring that application performance meets Service Level Agreements (SLAs), while preventing resource waste and performance degradation from improper resource over-commitment. Large-scale experimental validations conducted on the FC architecture within Huawei Cloud data centers demonstrate that, compared to EC, FC can reduce computing resource consumption by over 33% while managing the same workloads. Furthermore, FC's real-time performance degradation detection model achieves a prediction error of less than 5% across various testing environments, highlighting its commercial viability.

Realizing Generalized Zero-Shot Learning (GZSL) based on large models is emerging as a prevailing trend. However, most existing methods merely regard large models as black boxes, solely leveraging the features output by the final layer while disregarding potential performance enhancements from other layers. Indeed, numerous researchers have visually depicted variations in the features learned across different layers of neural networks. Motivated by this observation, we propose a Vision Transformer (ViT)-based GZSL method named Depth-Aware Multi-Modal ViT (DAM2ViT), which exploits multi-level features of ViT. DAM2ViT incorporates a multi-modal interaction block to align semantic information of categories across multiple layers, thereby augmenting the model's capacity to learn associations between visual and semantic spaces. Extensive experiments conducted on three benchmark datasets (i.e., CUB, SUN, AWA2) have showcased that DAM2ViT achieves competitive results compared to state-of-the-art methods.

Deep neural networks (DNNs) are widely used in fields like computer vision and natural language processing. A key component of DNN training is the optimizer. SGD-Momentum is popular in many DNN methodologies, such as ResNet and DenseNet, due to its simplicity and effectiveness. However, its slow convergence rate limits its use. To overcome this, we introduce inter-gradient collision into SGD-Momentum, inspired by the elastic collision model in physics. This new method, called ICSGD-Momentum, aims to improve convergence. We provide theoretical proof of convergence and establish a regret bound for ICSGD-Momentum. Experiments on benchmarks including function optimization, CIFAR-100, ImageNet, Penn Treebank, COCO, and YCB-Video show that ICSGD-Momentum accelerates training and enhances the generalization performance of DNNs compared to optimizers like SGD-Momentum, Adam, Radam, Adabound, and AdaBelief.

The advantages of resource elasticity and proactive data backup in cloud databases have attracted a large number of users to consider deploying their IT systems in the cloud. Factors such as performance, reliability, security, and ease of use are the most important considerations for users when choosing to migrate to the cloud. This paper mainly discusses the necessity of Quality of Service (QoS) awareness for cloud applications in the current public cloud environment, as well as the technological trends and challenges associated with it.

In cloud computing, tenants tend to select predefined flavors with substantial resources that go beyond their needs, in order to avoid resource starvation. As a result, cloud service providers experience a high allocation ratio with a low utilization ratio of resources. To improve the resource utilization, cloud service providers offer lower-priced resources by building an overcommitted environment to reuse resources. However, the colocation interference of workloads in the overcommitted environment may cause performance anomalies and degrade quality of service (QoS). In this paper, we propose a real-time mechanism to pin a group of workloads to a certain subset of CPUs to increase resource utilization and minimize interference. To be specific, we leverage deep reinforcement learning (DRL) to automatically adjust the CPU pinning configuration in the overcommitted environment according to the performance of workloads.

Breast cancer is the most commonly diagnosed cancer worldwide, and early detection is essential for reducing mortality rates. Digital mammography is currently the best standard for early detection, as it can assist physicians in treating the disease. However, inaccurate diagnoses from mammography are common and can lead to patients undergoing unnecessary tests and treatments. To address this challenge, deep-learning techniques have shown promising results in improving the accuracy and reliability of breast cancer detection. However, existing methods face two primary challenges: the lack of the annotated data, and the inability to adapt to new data domains. In this paper, we propose SelfAdaptNet to address these issues. Specifically, SelfAdaptNet employs self-supervised learning techniques, such as Bootstrap Your Own Latent (BYOL) and Simple Framework for Learning of Visual Representations (SimCLR), to tackle the problem of limited annotated data. Additionally, the adversarial technique is used to address the problem of domain shift. By successfully reducing domain disparities, this strategy enhances the model’s adaptability and robustness across a variety of clinical scenarios. Overall, our contributions offer a more effective and flexible approach for early breast cancer detection, and experimental results demonstrate that SelfAdaptNet can produce promising results as compared with other methods.

Partial domain adaptation (PDA) attempts to learn transferable models from a large-scale labeled source domain to a small unlabeled target domain with fewer classes, which has attracted a recent surge of interest in transfer learning. Most conventional PDA approaches endeavor to design delicate source weighting schemes by leveraging target predictions to align cross-domain distributions in the shared class space. Accordingly, two crucial issues are overlooked in these methods. First, target prediction is a double-edged sword, and inaccurate predictions will result in negative transfer inevitably. Second, not all target samples have equal transferability during the adaptation; thus, "ambiguous" target data predicted with high uncertainty should be paid more attentions. In this article, we propose a critical classes and samples discovering network (CSDN) to identify the most relevant source classes and critical target samples, such that more precise cross-domain alignment in the shared label space could be enforced by co-training two diverse classifiers. Specifically, during the training process, CSDN introduces an adaptive source class weighting scheme to select the most relevant classes dynamically. Meanwhile, based on the designed target ambiguous score, CSDN emphasizes more on ambiguous target samples with larger inconsistent predictions to enable fine-grained alignment. Taking a step further, the weighting schemes in CSDN can be easily coupled with other PDA and DA methods to further boost their performance, thereby demonstrating its flexibility. Extensive experiments verify that CSDN attains excellent results compared to state of the arts on four highly competitive benchmark datasets.

Traffic Sign Recognition (TSR) is an essential component of Intelligent Transportation Systems (ITS) and intelligent vehicles. TSR systems based on deep learning have grown in popularity in recent years. However, since these models belong to the closed-world-oriented learning paradigm, they are only capable of accurately identifying traffic signs that are easy to collect and cannot adapt to the real world. Furthermore, the sample utilization of these methods is insufficient, the resource consumption of model training may become unbearable as the data scale grows. To address this problem, we propose a novel “knowledge + data” co- driven solution (i.e., Joint Semantic Representation algorithm, JSR) for TSR. JSR creates a hybrid feature representation by extracting general and principal visual features from traffic sign images. It also realizes the model’s reasoning ability to zero-shot TSR based on prior knowledge of traffic sign design standards. The effectiveness of JSR is demonstrated by experiments on four benchmark datasets and two self-built TSR datasets.

Deep learning suffers from many notorious issues such as low convergence rate, over-fitting, and time-consuming. To alleviate these problems, an alternative learning framework with a non-iterative training mechanism named Broad Learning System (BLS) was proposed, which randomly assigns the parameters of hidden nodes and frozen them throughout the training process and then obtains its output weights using the ridge regression theory. This training method makes BLS have very high training efficiency. However, using ridge regression to solve the output weights cannot guarantee the stability of the solution in many cases, especially when the number of training samples is large, which may cause over-fitting and instability of BLS models. To solve this problem, we propose an improved BLS with a dense architecture and use the Proportional-Integral-Differential (PID) and Adaptive moment estimation (Adam) to replace the ridge regression operation. The new algorithm is called PID-A-DBLS, and its advantages include: 1) dense architecture can improve the feature extraction ability of the model; 2) using PID and Adam to solve the output weights can avoid the disadvantages of ridge regression. Extensive experimental results on four benchmark data sets show that PID-A-DBLS can achieve much better generalization ability and stability than BLS and its variants.

Conventional training mechanism for deep learning, which is based on gradient descent, suffers from many notorious issues such as low convergence rate, over-fitting, and time-consuming. To alleviate these problems, a novel deep learning algorithm with a different learning mechanism named Broad Learning System (BLS) was proposed by Prof. C. L. Philip Chen in 2017. BLS randomly selects the parameters of the feature nodes and enhancement nodes during its training process and uses the ridge regression theory to solve its output weights. BLS has been widely used in many fields because of its high efficiency. However, there is a fundamental problem that has not yet been solved, that is, the appropriate value of the parameter λ for the ridge regression operation of BLS is difficult to be set properly, which often leads to the problem of over-fitting and seriously limits the development of BLS. To solve this problem, we proposed a novel Dense BLS based on Conjugate Gradient (CG-DBLS) in this paper, in which each feature node is connected to other feature nodes and each enhancement node is connected to other enhancement nodes in a feed-forward fashion. The recursive least square method and conjugate gradient method are used to calculate the output weights of the feature nodes and enhancement nodes respectively. Experiment studies on four benchmark regression problems from UCI repository show that CG-DBLS can achieve much lower error and much higher stability than BLS and its variants.

Software vulnerabilities are usually caused by design flaws or implementation errors, which could be exploited to cause damage to the security of the system. At present, the most commonly used method for detecting software vulnerabilities is static analysis. Most of the related technologies work based on rules or code similarity (source code level) and rely on manually defined vulnerability features. However, these rules and vulnerability features are difficult to be defined and designed accurately, which makes static analysis face many challenges in practical applications. To alleviate this problem, some researchers have proposed to use neural networks that have the ability of automatic feature extraction to improve the intelligence of detection. However, there are many types of neural networks, and different data preprocessing methods will have a significant impact on model performance. It is a great challenge for engineers and researchers to choose a proper neural network and data preprocessing method for a given problem. To solve this problem, we have conducted extensive experiments to test the performance of the two most typical neural networks (i.e., Bi-LSTM and RVFL) with the two most classical data preprocessing methods (i.e., the vector representation and the program symbolization methods) on software vulnerability detection problems and obtained a series of interesting research conclusions, which can provide valuable guidelines for researchers and engineers. Specifically, we found that 1) the training speed of RVFL is always faster than Bi-LSTM, but the prediction accuracy of Bi-LSTM model is higher than RVFL; 2) using doc2vec for vector representation can make the model have faster training speed and generalization ability than using word2vec; and 3) multi-level symbolization is helpful to improve the precision of neural network models.

The initialization of neural networks with random weights (NNRW) has a significant impact on model performance. However, there is no suitable way to solve this problem so far. In this paper, the relationship between meta features of a dataset and the initialization of NNRW is studied. Specifically, we construct seven regression datasets with known attributes' distributions, then initialize NNRW with different distributions and trained them based on the datasets to get the corresponding models respectively. The relationship between the attributes' distributions of the datasets and the initialization of NNRW is analyzed by the performance of the models. Several interesting phenomena are observed: firstly, initializing NNRW with the Gaussian distribution can help the model to have a faster convergence rate than ones with the Gamma and Uniform distribution. Secondly, if one or more attributes in a dataset that follow the Gamma distribution, using Gamma distribution to initialize NNRW may result in a slower convergence rate and easy overfitting. Thirdly, initializing NNRW with a specific distribution with smaller variances can always achieve faster convergence rate and better generalization performance than the one with larger variances. The above experimental results are not sensitive to the activation function and the type of NNRW. Some theoretical analyses about the above observations are also given in the study.