Vol. 44 · No. 5

Serverless computing brings another revolution to cloud computing as function-as-a-service (FaaS), where the applications are abstracted as a group of functions. Serverless applications are cost-effective and manage resources efficiently, but the lack of performance modeling and energy optimization affects the potential users' broad adoption of serverless computing. Performance enhancement and energy optimization are necessary to guarantee serverless applications' service level agreement (SLA). This review paper presents various performance metrics in serverless computing, including cost, scalability, latency, energy consumption, resource utilization, fault tolerance, and response time. Based on these metrics, various performance modeling and energy optimization techniques have been explored to reduce energy consumption and improve system efficiency. Furthermore, the review investigates software platforms for implementing serverless computing, including AWS Lambda, Apache OpenWhisk, Azure Functions, and Google Cloud Functions, highlighting key findings and limitations. This comprehensive review serves as a guide for researchers, directing them toward new and promising research directions in the field.

Graph Neural Networks (GNNs) have shown great promise in functional Magnetic Resonance Imaging (fMRI) analysis due to their ability to capture complex interactions between brain regions. However, existing models often overlook the brain's physiological structure and fail to leverage hierarchical information from brain atlases. In this paper, we propose Hierarchical Bilateral Graph Neural Network (HiBiGNN), a generic architecture that integrates hierarchical information from brain atlases and incorporates the bilateral structure of the brain, with the ability to be instantiated with various existing GNNs as its foundation. HiBiGNN processes a special heterogeneous graph structure, called the Hierarchical Bilateral Graph (HiBiG), which combines multi-level brain graphs derived from functional regions defined by multi-level brain atlases and divides each brain graph into left and right subgraphs, thereby modeling multiple types of nodes and relations. During feature extraction, HiBiGNN performs deep fusion of features from different types of nodes using a unique convolution operation (HiBiG-Conv) and generates graph-level representations via a specialized readout operation (HiBiG-Readout) for graph classification tasks. To assess the effectiveness of HiBiGNN, we conducted extensive experiments on a graph classification task using an fMRI dataset we collected from a response inhibition task, testing multiple HiBiGNN instances with different base GNN models. The results show that our HiBiGNN instances outperforms several generic GNN models as well as those specifically designed for fMRI analysis, demonstrating the significant potential of HiBiGNN for future applications.

Identifying a person's age and gender from speech signal characteristics poses a significant challenge in personal identity recognition systems, particularly when security considerations are involved. In signal processing applications such as speaker recognition, biometric identification, human-machine interface (HMI), and telecommunication, the estimation of age and gender from voice is a crucial and demanding problem. In several signal processing domains, deep learning models have demonstrated remarkable effectiveness. In this paper, we propose a modified convolutional neural network to identify the age and gender of the speaker using the characteristics of the MFCC speech. We also included techniques to reduce the dimensionality of the speech feature set. We tested modified one-dimensional convolutional neural networks (1D-CNN) and machine learning models such as support vector classification (SVC), decision trees (DT), and random forests (RF). The modified 1D-CNN based on deep learning, along with dimensionality reduction, random seeding, and cross-validation, is proposed for the recognition of age and gender in speech. We applied different dimensionality reduction techniques, such as principal component analysis (PCA) and independent component analysis (ICA), along with random seeding and various sets of cross-validation. In this study, we used the Children Speech Recorning Dataset, Biometric Visions and Computing (BVC), and the Mozilla Common Voice speech datasets for estimating age and gender from speech. The proposed 1D-CNN model exhibits a promising performance compared to the state-of-the-art (SOTA) approaches. The models were evaluated and compared with evaluation metrics, such as accuracy. The dimensionality reduction techniques, selection of speech features, and seeding show a significant impact on the performance of the suggested model.

A stroke is a serious neurological condition that occurs due to either blockages or bleeding in the brain, which can lead to death or long-term disability. This study aims to enhance the accuracy of disease diagnosis in imbalanced stroke patient datasets. The model incorporates an artificial immune system algorithm, whose parameters are fine-tuned using the Firefly algorithm to ensure both stability and balanced data. To enhance the performance of the underrepresented class, the One-Sided Selection method is employed. The model's effectiveness was tested in two separate experiments: one utilizing all available features and the other applying the Artificial Bee Colony (ABC) algorithm to select the most relevant features. The models were trained using six different classification algorithms: CatBoost, Light Gradient Boosting Machine (LightGBM), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Logistic Regression (LR). The results were presented using performance metrics. When trained using all features, the model achieved an accuracy of 93

Current multi-modal emotion recognition techniques primarily use modalities such as expression, speech, text, and gesture. Existing methods only capture emotion from the current moment in a picture or video, neglecting the influence of time and past experiences on human emotion. Expanding the temporal scope can provide more clues for emotion recognition. To address this, constructed the Situation-Based Video Context Emotion Datasets (SVCEmotion) dataset in video form. Experiments show that both VGGish and BERTbase achieve good results on SVCEmotion. Comparison with other audio emotion recognition methods proves that VGGish is more suitable for audio emotion feature extraction on the dataset constructed in this paper. Comparison experiments with textual descriptions demonstrate that the contextual descriptions introduced in the SVCEmotion dataset for the emotion recognition task under wide time range can provide clues for emotion recognition, and that the combination with factual descriptions can substantially improve the emotion recognition effect.

Large language models (LLMs) aid programmers in understanding code but are limited by input length when handling large codebases. To address this, we propose using Goal Structuring Notation (GSN) – originally developed for articulating assurance cases in complex engineering projects – to represent and break down large codebases. We introduce a tool that leverages LLMs to automatically convert large code into GSN. The generated GSN provides an overview that simplifies code comprehension and enhances communication among programmers. Experimental results demonstrate that our approach significantly increases programmers' confidence levels and reduces task completion times.

Clustering analysis is a crucial method in data mining, aimed at identifying clusters of data objects in the attribute space. Distributed clustering has gained prominence due to the emergence of Big Data. The rapid growth of data, particularly with the advent of technologies, such as the Internet of Things and 5G, has resulted in numerous challenges for data analysis and processing. Traditional clustering methods, such as K-means and hierarchical clustering, struggle with extensive datasets designed for smaller to moderately sized datasets. Meta-heuristic techniques have garnered significant attention among the various distributed clustering algorithms due to their ability to deliver high-quality solutions across a wide range of optimization problems. In this study, we proposed a new Cuckoo search (CS) clustering algorithm for distributed clustering to address the challenges of Big Data clustering. First, the CS clustering algorithm is executed on each local site, utilizing GPU acceleration for efficient local data clustering. Second, on a global scale, representative data from each site are aggregated and processed worldwide, with centroids iteratively updated to generate the final clustering result. We have significantly enhanced the processing efficiency by minimizing transmission costs and eliminating the need for inter-node communication. Furthermore, our approach demonstrates adaptability in handling large datasets with competitive execution times through the utilization of parallel processing and distributed computing. Our approach demonstrates both efficiency and scalability across wide range of datasets, highlighting its potential for various applications.

Catastrophic forgetting is a major challenge for online class-incremental learning. Existing replay-based methods have achieved a certain degree of effectiveness, but are limited by not considering the quality of the samples and the key semantic information in a single-pass data stream. To address these issues, we proposed the framework of Online Class-Incremental Learning Based on Attention Distillation(ADOCIL), which consists of three parts. A two-stage sampling method is used in the replay stage to improve the quality of the samples taken. Meanwhile, we introduced the Attention-based Dual-View Consistency (ADVC), which enables the model to fully explore the critical semantic information within a single-pass data stream. In addition, to further mitigate the problem of catastrophic forgetting, we introduced attention distillation to map the attentional map of the teacher model to the student model, thus solving the problem of forgetting historical tasks. Extensive experiments demonstrated the effectiveness of ADOCIL.

Weibo has emerged as a vital platform for Chinese netizens to share information, but it has also given rise to numerous rumors. Real-time detection methods that do not rely on propagation features are the most effective way to curb the spread of these rumors. Currently, real-time detection methods that mine semantic features of rumor text based on deep learning lack sufficient generalization ability. Therefore, we propose a real-time rumor detection method integrating multiple user and content features. In addition to standard user basic features, our approach utilizes the user's historical posting data to extract two deep-level features: user rationality and professionalism. Regarding content features, in addition to standard statistical features, we use a graph attention network that considers edge weights to learn deep semantic features of the content. The user and content features are concatenated and fed into a multi-layer perceptron for classification. The experimental results on a real Weibo dataset show that the accuracy of the proposed method achieves 92.6

We discuss some progress in the design and implementation of block-Jacobi SVD algorithms in the scope of the last 25 years. Two ideas were crucial for enhancing the efficiency of two-sided and one-sided serial or parallel block-Jacobi methods: the so-called dynamic ordering of subproblems solved in each iteration step, and a suitable preconditioning of an original matrix. These two ideas led to a substantial decrease of (serial or parallel) iteration steps needed for the convergence. Consequently, especially the one-sided block-Jacobi SVD algorithm became competitive in speed with some SVD algorithms based on the matrix bi-diagonalization. We also discuss new theoretical results w.r.t. the asymptotic quadratic convergence of block-Jacobi SVD algorithms regardless of the distribution of singular values of the original matrix.