Federated learning is a distributed learning approach that balances data privacy and collaborative learning. To address the impact of non-IID (non-Independently and Identically Distributed) data on the performance of the global model, this study delved into personalized federated learning. First, we introduced a local network with a self-attention mechanism, enhancing the personalization of local models. Next, we proposed a weighted average aggregation method, assigning weights based on the distance of the client model parameters and variations in contributions, mitigating the adverse effects of uneven data distribution. Using a dataset constructed from non-IID data partitions simulated via the Dirichlet distribution, we experimented, integrating the self-attention mechanism of the local model and the personalized aggregation strategy. Experimental results indicate that this approach significantly improves the test accuracy of the MNIST and SVHN datasets while reducing generalization error. In addition, it also reduces the oscillation during training, showing good robustness. In conclusion, this study offers fresh insights into handling non-IID data in federated learning and empirically validates the method’s effectiveness.
Multimodal sentiment analysis tasks have been extensively researched in recent years. However, real-world multimodal data collection is often problematic due to missing modalities, making it suboptimal. To address these issues, this paper presents a supervised contrast learning and cross-modal translation-based multimodal sentiment analysis model that can operate even when modalities are missing. First, supervised contrast learning is employed to carry out cluster analysis on the same sample batch and establish the connection between different samples. Incomplete features of the target samples are then supplemented with the other samples. Following that, a cross-modal translation network, based on Transformer, is utilised to learn the collective representation of different modalities. A feature reconstruction network is used, employing center distance difference loss, to reduce the difference between features obtained by the model under modality-missing and modality-complete conditions. This aims to enhance the robustness of feature extraction. Experimental findings from two publicly available datasets demonstrate the model’s improved performance for varying levels of modalities missing.
This paper introduces a novel approximate multiplier tailored for FPGAs, with a primary focus on its application in hardware accelerators for deep machine learning computations. In deep learning, multipliers represent a substantial portion of the hardware complexity. The primary objective is to enhance hardware efficiency compared to exact multipliers while maintaining an acceptable level of accuracy, especially for inference tasks. The proposed multiplier employs a parallel architecture featuring a configurable set of approximate adders. The paper provides comprehensive insights into the design of both INT8 and UINT8 configurations of this multiplier. The evaluations encompass error analysis findings, hardware implementation results, and accuracy measurements for the inference of a selected set of benchmark deep learning models. The results illustrate that a chosen configuration of this multiplier can achieve a notable 9% reduction in overall LUT utilization while incurring only an average 0.28% and 0.06% reduction in inference accuracy before and after approximation-aware retraining, respectively.
Parked vehicle assisted edge computing is effect to alleviate the resource supply-demand conflict in vehicular edge computing, by leveraging the idle resources in parked vehicles (PVs). The uncertain parking behavior and the load balance both of PVs are the critical factors to the quality-of-service (QoS) for users, which have not been well jointly addressed. To fill this gap, this paper formulates a load balance problem with the goal of minimizing the maximum load among PVs, under the constraints of service reliability and QoS for tasks. The NP-hardness of the problem is proved. Due to the NP-hardness and coupled variables, the problem is decomposed into two subproblems, i.e., task offloading problem and resource allocation problem. An efficient algorithm is proposed to iteratively solve the two subproblems. For the first subproblem, this paper derives a lower bound of the allocated computing resources for each task, and designs a greedy algorithm to make offloading decision based on the derived bound. For the second subproblem, this paper proves its convexity and proposes a subgradient algorithm to make resource allocation. The service reliability of PVs is evaluated from the traces in a real-world dataset. Experimental results show that, the proposed algorithm can effectively reduce the maximum load among PVs, compared with the existing algorithms.
code_saturne is an open-source general large-scale computational fluid dynamics software independently developed by Electricite de France since 1997. Due to its high computing performance, code_saturne has been widely used in the nuclear industry, heavy fuel and biomass combustion, particle tracking and other computational fluid dynamics fields. In this work, we implemented code_saturne on the new generation Sunway supercomputer (SW26010pro) at the first time. By applying the data value-taking technology of double cache, combined with Sunway shared memory technology, we have adapted the data structure and Modular programming of code_saturne conducively to the new generation Sunway supercomputer (SW26010pro) and accelerated the software hotspots the software hotspots by 1–2 times. After optimizing and porting, this software has successfully served in the numerical simulation of the lateral flow of fuel assemblies in the nuclear industry.
The memory hierarchy comprises main memory, which can store a large amount of data, and cache memory, which can access data at high speeds. By accessing data stored in the same array continuously, we can use cache memory without performing frequent data replacement from the main memory, which is crucial for high-speed execution of programs. In this study, we propose a novel code motion algorithm, named global store statement aggregation (GSA), to improve program execution by efficiently utilizing the memory hierarchy. To achieve this, GSA moves each store statement referring to an array immediately before the following store statements accessing the same array. We implemented GSA in a real compiler and evaluated it using sorting programs. The experimental results indicate that our algorithm effectively reduces the number of cache misses in comparison to a previous code motion algorithm.
In Vertical Federated Learning, Split Learning has gained popularity as a technique where a neural network is split into two parts held by two parties. One party holds the input data, and another party holds the corresponding labels. The two parties exchange intermediate features and gradients of features to update the network parameters to preserve the privacy of the private data and network parameters. However, recent studies just utilized a small amount of auxiliary data to infer private labels. In the paper, we propose a label inference attack method based on soft label by leveraging the auxiliary data. To further enhance the performance of the attack method, we design a loss function that combines a gradient-matching loss and a model loss based on soft label. Our experimental evaluation is conducted on classification datasets, and the results demonstrate that our model has superior attack effectiveness and generalization capabilities.
In response to the numerous challenges present in the current management of railway engineering graphical information, including poor data collaboration among project stakeholders, low information flow, and insufficient correlation between 2D drawings and 3D models, this paper integrates the characteristics of abundant data volume and multidisciplinary aspects in railway engineering graphical information. Drawing upon theoretical foundations such as digital model separation, model lightweighting, and database languages, along with leveraging technologies such as Web Graphics Library (WebGL) - based 3D engines and specialized Engineering Breakdown Structure (EBS) encoding tailored to railway engineering, the paper presents an integrated approach for managing two- and three-dimensional models in railway engineering based on a unified graphical model. This approach is rooted in the use of Globally Unique Identifier (GUID) to tightly associate 2D drawings and 3D models, while also standardizing data formats and sharing protocols. As a result, it significantly enhances the efficiency of project management and collaboration within engineering construction projects, effectively breaking down data silos. The method has been successfully applied to a railway engineering project, offering a reliable and innovative solution for managing its information deliverables.
GigaPixel-level computer vision tasks recently become new research hotspots, due to the development of photography. Object detection, as a basic, common, but challenging task, undoubtedly received the most attention. However, most of the research focused on the efficiency improvements for the super resolution of the scenarios. They tend to design relevant network modules or inference strategies to help split the whole image into smaller patches for efficient computation. Differently from them, We proposed three optimization strategies that can maintain efficient computation while also ensuring the accuracy of model inference strategies. The strategies are Anchor-Split Sample Strategy, GPU Memory Optimizations and Two-Phase Adaptive Inference Strategy. Anchor-Split Sample Strategy can help train the detectors within 8 hours on the PANDA detection datasets. GPU Memory Optimizations can help train the DETA model with Swin-Large backone on a consumer GPU card like RTX 3080 with only costing 18G memory. Two-Phase Adaptive Inference Strategy, Without the extra training of the additional network modules or complex strategies, can obtain 74% mAP and 82% AR500 performance with only 1.5h cost on the 15W Power Jetson Orin AGX card. Compared with the state-of-the-art methods, our methods can boost the performance by 20% percentage.
Federated learning is a distributed machine learning framework that allows multiple clients to collaborate in training a global model. In federated learning, many clients can train together and make full use of information resources. However, in practical scenarios, the data owned by different users are not from the same distribution. The integral data is heterogeneous, which will slow down the training speed of federated learning and reduce the performance of the model. And in this paper, we focus on addressing label heterogeneity and propose a clustering federated learning framework that utilizes the flow of solution procedure (FSP) matrix. And we set up a small amount of public dataset at the central server. The client participating in the training generates FSP matrix on this dataset. The central server clusters the clients into several groups according to these matrixes. The clients clustered into the same group are considered to have more similar data distributions. We also conduct quantities of experiments on CIFAR10 and MNIST datasets to verify that our method performs better than the traditional federated learning and some other algorithms in terms of performance and has good stability. It can function properly in various scenarios with non-independent and non-identically distributed data.
To solve the Straggler problem which seriously affects computational efficiency in distributed gradient descent (DGD), we design a dynamic redundant gradient coding method(DRGC) based on Reed-Solomon coding with dynamic code words and heartbeat delay-aware mechanism that can monitor the computational progress of workers. Traditional coding method ensures the straggler tolerance through a large of redundant computations, but can’t adapt to the change of cluster state. In contrast, DRGC can flexibly adapt to the change by changing the code words base on the cluster state that heartbeat delay-aware mechanism has monitored and determined. Experiments prove that compared with the RS gradient coding, DRGC can further reduce training time.
The dragonfly topology has been widely utilized in large-scale interconnection networks due to its scalability and low diameter. However, a large dragonfly system requires a large number of high-radix routers, resulting in high hardware costs. This paper introduces Recursive Dragonfly (RDF) as a solution to overcome this problem. RDF uses a new parameter, n, which represents the number of recurrences in the network. The RDF is constructed based on a recursive approach. For n = 1, RDF acts the same as a traditional dragonfly. For n = 2, each traditional dragonfly network forms a group within the RDF network. In general, for RDF with n ≥ 2, each group can be viewed as an RDF network with n − 1. The new parameter n allows RDF to construct an interconnection network with various diameters and radices. Our evaluation results show that RDF achieves lower hardware costs while maintaining shorter packet latencies compared to the traditional dragonfly network. RDF topology is more suitable for constructing large-scale interconnection networks.
As the infrastructure for the development of digital economy, the Internet of Things (IoT) is widely used in electric power, transportation, industry and other fields, bringing digital services to the government, enterprises and individual consumers. At the same time, the IoT also carries more and more production and life information, and its security issues can not be ignored. Besides, some malicious attacks will affect the industry security. Compared with the traditional Internet, the IoT has the characteristics of multi-source, heterogeneity, openness, ubiquity, etc. Due to the defects in the authenticity and reliability of the data itself, the single-dimensional security protection technologies such as data encryption, identity authentication and anomaly detection cannot provide the whole process and system-level trustiness guarantee, which makes it face a more complicated and severe trust crisis. In this paper, the requirements and characteristics of the trusted IoT are firstly introduced. Secondly, the challenges faced by the trusted IoT are analyzed and after the corresponding solutions are proposed. Finally, the development suggestions of the trusted IoT are elaborated.
Focus on the problem of long loading time of complex model and user QOE in WebBIM scene, this paper proposes a hierarchical model loading strategy in WebBIM scene based on cloud-edge-terminal collaborative architecture. First of all, with edge computing technology, we apply the cloud-edge-terminal collaborative architecture to BIM scenarios to speed up model loading and service response processes. Next, by following the architectural semantics, the transformed IFC file is divided into index, geometry, material and related attributes, and component priorities are designed using geometry, semantics and user relevance to realize model loading based on the cloud-edge-terminal collaborative architecture. Finally, a cloud-edge-terminal collaborative BIM experimental scene is built, and we select 7 IFC scenes to compare initial loading time, earliest operable time, and effective pixel filling degree with BIMServer, the experimental results show that our strategy can effectively improve the above three aspects of performance and user QOE in WebBIM scene.
An empathic dialogue generation model based on comparative learning and multi-source data fusion is proposed to address problems of the under-extraction of contextual emotional expressions in discourse and limited empathic response diversity in current empathic dialogue generation tasks. The model introduces cognitive, affective, and dialogue intention information to guide dialogue generation. In the emotion classification stage, contrastive learning is used to capture the emotion information of the contextual discourse and to improve the contextual discourse representation. In the empathic response generation phase, the multi-attention mechanism fuses multiple sources of cognitive, emotional, and dialogue intention information, capturing more in-depth semantic and emotional information in the dialogue. Experimental results on the public dialogue dataset ED show that the proposed model can fully exploit cognitive, affective, and dialogue intent information to improve classification accuracy and generate more empathetic dialogue responses.
At present, China railway development is in a critical period of strategic opportunities. Smart railway have become an important development direction. Digital twins is an important mean for smart railway construction, which can solve the problem of information and physical integration and realize the application of smart railway. Therefore, based on the analysis of the relationship between digital twins and smart railway, this paper proposes a railway digital twin model based on the five-dimensional digital twin model, and explores the description framework of the railway twin system and the railway digital twin architecture, which provides a certain reference for the application of digital twins in the railway field.
In recent years, graph matching has made substantial progress in planar object tracking, offering end-to-end learning capabilities. However, these methods often fall short in real-world tracking tasks due to their reliance on static, manually constructed graphs that fail to effectively account for the dynamic distribution of feature points and their relationships within individual video frames. To address these challenges, we introduce the adaptive graph matching (AGM) Network, a resilient object tracking model designed to handle complex transformations and changes in object appearance with significant accuracy. The cornerstone of AGM lies in its innovative construction of dynamic graphs that model the feature points across video frames. Nodes represent key feature points in the graph, while edges adaptively form based on the spatial relationships and descriptors of these points. This allows the graph to adjust flexibly to the properties and relationships inherent in the feature points of each frame, enhancing its fidelity to the actual structure and changes of the planar object across the video sequence. AGM employs self-attention mechanisms to guide the adaptive graph construction, and cross-attention mechanisms to capture the relationships between adjacency matrices and their associated features. Our ablation study shows that these self and cross-attention mechanisms play an integral role in enhancing the model’s tracking performance, providing a more robust solution to real-world complications such as motion blur and object occlusion. We trained AGM and validated its performance on three datasets. The results show that AGM consistently outperforms other state-of-the-art models, demonstrating the effectiveness of our approach.
The advancement of machine learning is significantly influenced by the accessibility of vast data. However, in reality, data is distributed across various entities and is subject to various legal and practical constraints. Therefore, it is unrealistic to merge these dispersed datasets into a single large dataset. In order to address this challenge, a novel framework called Federated Transfer Learning (FTL) has been introduced. FTL allows knowledge sharing without jeopardizing user privacy and enables the transfer of complementary knowledge across different domains. This capability enables the use of rich labels from the source domain to construct an effective model tailored to the target domain’s specific needs. In this paper, we propose an improved algorithm, named FTL-FIC, based on FTL for image classification datasets. It leverages federated learning to aggregate data and utilizes transfer learning to construct an shared learning model, thereby adapting to the characteristics of the target dataset. Experimental results demonstrate the algorithm’s notable improvement in the classification accuracy of the target dataset (increased by 8.17%), the reduced number of convergence epochs, and its ability to maintain good generalization on the source dataset. Furthermore, FTL-FIC exhibits versatility and scalability, indicating its potential application to other types of datasets.
In this paper, we design a fine-grained task execution scheme under the cooperation of terminals, edge and cloud servers. Firstly, we build a virtual hierarchical architecture for terminal-edge-cloud cooperative network and introduce SDN technology to balance the edge load and achieve efficient network information management. Then our task execution scheme is proposed with two task serving patterns and three execution mechanisms, specifically, for tasks with pattern of “served as a whole”, we build an optimization problem to minimize delay and energy consumption jointly; for tasks with pattern of “served with segmentations”, a Subtask Scheduling Parallelism Maximization (SSPM) algorithm based on dynamic list scheduling is designed to maximize the parallelism degree of subtasks execution as much as possible. Finally, we carry out our simulation on NS-3 platform.
In the era of information explosion, people are constantly exposed to a wealth of unreliable information. Nevertheless, societal stability and public trust may be seriously threatened by such unverified or even fabricated false information. In order to discriminate the fake news, several fake news detection approaches have been proposed. However, most of the current multimodal fake news detection methods only rely on news text and images from a single data source. Owing to the scarcity of multimodal information and the limited volume of data, the effectiveness of the majority of these methods is low. As a result, we embrace the approach of amalgamating multi-source heterogeneous data and propose a new network structure FusionNet. In this network structure, we creatively use web crawler technology to expand the data. We achieve double amplification of text and image as opposed to the prior single-mode amplification of text, and by data screening and cleaning, we create multi-source and multimodal datasets for various viewpoints on the same event. We also propose a novel modal fusion approach that leverages the inherent information in news content, integrates an attention mechanism, and explores the shared relevant features between original news and multi-source news. Through empirical results and analysis, our approach is superior to other methods.
