Computer Engineering & Science

A survey of error correction codes in holographic storage

YU Qin, Wu Fei, ZHANG Meng, XIE Chang-sheng

2024, 46(04): 571-579. doi:

Abstract ( 114 )

PDF (1981KB) ( 297 ) 　　

In the era of big data, the demand for high-density and large-capacity storage technology is increasing day by day. Unlike traditional storage technologies that record data bit by bit, holographic storage uses two-dimensional data pages as the unit for reading and writing, adopting a three-dimensional volume storage mode. With advantages such as high storage density, fast data conversion rate, energy efficiency, safety, and ultra-long-term preservation, holographic storage is expected to become the strong competitor for mass cold data storage. This paper focuses on phase-modulated collinear holographic storage and analyzes the current research status of error correction codes for holographic storage. A detailed introduction is provided to an reference beam-assisted low-density parity-check (LDPC) code scheme.

Convolutional neural network inference and training vectorization method for multicore vector accelerators

CHEN Jie, LI Cheng, LIU Zhong

2024, 46(04): 580-589. doi:

Abstract ( 93 )

PDF (982KB) ( 292 ) 　　

With the widespread application of deep learning, represented by convolutional neural networks (CNNs), the computational requirements of neural network models have increased rapidly, driving the development of deep learning accelerators. The research focus has shifted to how to accelerate and optimize the performance of neural network models based on the architectural characteristics of accelerators. For the VGG network model inference and training algorithms on the independently designed multi core vector accelerator FT-M7004, vectorized mapping methods for core operators such as convolution, pooling, and fully connected layers are proposed. Optimization strategies, including SIMD vectorization, DMA double-buffered transfer, and weight sharing, are employed to fully exploit the architectural advantages of the vector accelerator, achieving high computational efficiency. Experimental results indicate that on the FT-M7004 platform, the average computational efficiency for convolution layer inference and training is 86.62% and 69.63%, respectively; for fully connected layer inference and training, the average computational efficiency reaches 93.17% and 81.98%, respectively. The inference computational efficiency of the VGG network model on FT-M7004 exceeds that on the GPU platform by over 20%.

Feature extraction and prediction of multidimensional time series based on GGInformer model

REN Sheng-qi, SONG Wei

2024, 46(04): 590-598. doi:

Abstract ( 91 )

PDF (740KB) ( 207 ) 　　

With the rapid development of big data and Internet of Things (IoT) technologies, the application scope of multidimensional time series data has become increasingly widespread. Faced with a large amount of complex time series data characterized by non-linearity and high-dimensional redundant features, traditional time series analysis methods struggle to effectively address the complexity of multidimensional time series with high-dimensional features, resulting in suboptimal predictive performance. To address these issues, this paper proposes the GGInformer model, which improves upon the Genetic Algorithm and Informer model while incorporating the GRU network. This model not only efficiently extracts key features from multidimensional time series but also effectively addresses long-term dependency issues. To validate the predictive capability of the model, experiments are conducted on two real datasets and three public benchmark datasets, all of which demonstrated superior performance compared to the baseline models. Specifically, compared to the Informer baseline model, the GGInformer model achieves reductions in Mean Squared Error (MSE) values of 22%, 13%, 20%, 23%, and 38% across the five datasets. The experimental results indicate that the GGInformer model can effectively address the complex feature extraction challenges of multidimensional time series data and further enhance time series prediction capabilities.

A low-power transmitter driver for die to die

REN Bo-lin, XIAO Li-quan, QI Xing-yun, ZHANG Geng, WANG Qiang, LUO Zhang, PANG Zheng-bin, XU Jia-qing

2024, 46(04): 599-605. doi:

Abstract ( 75 )

PDF (1091KB) ( 178 ) 　　

A low-power transmitter driver for chiplet interconnection was designed and experimentally implemented based on the inter-chip interconnection standard proposed by the UCIe protocol. The driver circuit adopts a source series terminated (SST) driver, whose power consumption is only 1/4 that of the current mode logic (CML) structure. In addition, based on adjustable feedforward equalization technology, the driver circuit adjusts the equalization strength for different channel attenuations. By de-emphasizing equalization, it enhances the quality of the transmitted signal, ultimately reducing inter-symbol interference. This circuit was designed under CMOS 28 nm process. The front-end simulation results show that the maximum equalization intensity is -3.7 dB when the 0.9 V voltage is supplied. When the 32 Gbps NRZ signal passes through the 21 mm channel (the attenuation at the 16 GHz Nyquist frequency is -2.37 dB), after adjusting the appropriate equalization intensity, the eye height of the output waveform eye diagram is 253 mV (71.8%), the eye width is 27 ps (87%), and the simulation power consumption is only 4.0 mW.

Hardware design and FPGA implementation of a variable pipeline stage SM4 encryption and decryption algorithm

ZHU Qi-jin, CHEN Xiao-wen, LU Jian-zhuang,

2024, 46(04): 606-614. doi:

Abstract ( 85 )

PDF (1475KB) ( 244 ) 　　

As the first commercial cryptographic algorithm in China, SM4 algorithm is widely used in data encryption storage and information encryption communication and other fields due to its advantages of simple and easy implementation of algorithm structure, fast encryption and decryption speed and high security. With the variable pipeline stage SM4 encryption and decryption algorithm hardware design and FPGA implementation as the research topic, this study focuses on the performance differences in designs with different pipeline stages. A controllable pipeline stage SM4 encryption and decryption circuit is designed and encapsulated into an IP core with AXI and APB interfaces. Based on XILINX ZYNQ devices, a small SoC is constructed on the XILINX ZYNQ-7020 development board, and the designed SM4 IP core is mounted onto the AXI bus for simulating real-world scenarios and conducting performance tests. The correctness of the design functionality is verified by comparing software encryption and decryption data with simulated data. Testing the performance of different pipeline stages helps identify the most suitable pipeline stage number.

Research on parallel acceleration of line cloud privacy attack algorithm

GUO Chen-liang, YAN Shao-hong, ZONG Chen-qi

2024, 46(04): 615-625. doi:

Abstract ( 56 )

PDF (1529KB) ( 176 ) 　　

The localization methods based on line cloud can protect scene privacy, but they also face the risk of being cracked by a privacy attack algorithm proposed by Kunal Chelani et al. This attack algorithm can recover approximate point clouds from line clouds, but its computational efficiency is low. To address this issue, a parallel optimization algorithm is proposed and evaluated in terms of running time and speedup ratio. Specifically, the CPU multi-core parallelism and the GPGPU parallelism are implemented using the SPMD pattern and the pipeline parallel pattern respectively. Furthermore, the data parallel pattern is adopted to implement heterogeneous computing, to achieve the highest degree of parallelism. Experimental results demonstrate that the maximum speedup ratio of the parallel optimization algorithm is 15.11, and the minimum is 8.20. Additionally, compared to the original algorithm, the parrellel optimization algorithm ensures the relative error of the recovered point clouds within 0.4% of the original error, ensuring the accuracy of the algorithm. This research holds significant importance and reference value for line cloud privacy attack algorithms, as well as for privacy protection algorithms in Line Cloud under different scenarios and other density estimation problems.

An optimal energy efficiency strategy based on mmWave cooperative communication small cell under SWIPT

LU Ming-yu, LI Tao-shen, L Pin

2024, 46(04): 626-634. doi:

Abstract ( 45 )

PDF (734KB) ( 163 ) 　　

Addressing the optimization problem during the simultaneous energy and information transmission stage in the simultaneous wireless information and power transfer (SWIPT), an optimal energy efficiency strategy based on mmWave cooperative communication small cell is proposed to maximize the link energy efficiency. This strategy, under the joint constraints of minimum link transmission rate and minimum collected energy, employs a power splitting mode at the receiving end of energy- constrained user devices. By optimizing transmit power control and power splitting factors, it aims to maximize the system's link energy efficiency. Recognizing the original problem as a non-convex fractional programming problem with NP-hard characteristics, the Dinkelbach method is utilized to transform the objective function into a convex optimization problem that is easier to solve. An iterative algorithm is designed to find the optimal solution through cross iterations. Simulation results demonstrate that the proposed strategy outperforms traditional power control methods and maximum transmit power approaches in optimizing system energy efficiency.

Semi-supervised website topic classification based on hetero-geneous graph neural networkWANG

Xie-zhong, CHEN Xu, JING Yong-jun, WANG Shu-yang

2024, 46(04): 635-646. doi:

Abstract ( 71 )

PDF (2434KB) ( 205 ) 　　

The rapid growth of the number of Internet websites has made existing methods challenging to accurately classify specific website topics. URL-based methods, for example, struggle to handle topic information not reflected in the URL, while content-based methods face limitations due to data sparsity and challenges in capturing semantic relationships. To address this, a semi-supervised website topic classification method, HGNN-SWT, based on a heterogeneous graph neural network, is proposed. This method not only utilizes website text features to complement the limitations of using only URL features but also models sparse relationships between website text and words using a heterogeneous graph, improving classification performance by handling node and edge relationships within the graph. The approach introduces a neighbor node sampling method based on random walks, considering both local features and the global graph structure of nodes. Additionally, a feature fusion strategy is proposed to capture contextual relationships and feature interactions within website text data. Experimental results on a self-created Chinaz Website dataset demonstrate that HGNN-SWT achieves higher accuracy in website topic classification compared to existing methods.

A dual-verification model watermarking scheme based on certification files

WU Xia, ZHENG Hong-ying, XIAO Di

2024, 46(04): 647-656. doi:

Abstract ( 60 )

PDF (1037KB) ( 158 ) 　　

With the integration of edge computing frameworks and federated learning protocols, an increasing number of copyright protection methods for deep learning models have been proposed. However, solely verifying ownership from the senders perspective does not provide assistance to the receiver. Numerous studies have indicated that in client-edge-cloud federated learning systems, malicious users attempt to gain access to public models without contributing or even poison the public models. Therefore, it is necessary to provide a model ownership verification scheme for the receiver. Building upon existing neural network watermarking schemes, this paper proposes a dual-verification model watermarking scheme based on certification files. It introduces a certification file generation step and implements dual ownership verification of the model through adjustments in the model structure. The feasibility, robustness, and improvement in watermark embedding rate of the scheme are verified through experiments.

A color image encryption method based on improved 3D_Henon chaos mapping

NIU Shi-ming, XUE Ru, DING Cong

2024, 46(04): 657-666. doi:

Abstract ( 70 )

PDF (2223KB) ( 163 ) 　　

Aiming at the problems of small chaotic space and weak chaotic ability of existing chaotic mapping models, a new three-dimensional chaotic mapping model is designed by coupling two- dimensional Henon chaotic mapping model and Sine chaotic mapping model. Aiming at the problem that the security of encryption system will be reduced if the key is reused, this paper designs a new key generation method using Chebyshev mapping and plaintext image. Using the 3D chaotic mapping and key generation method designed in this paper, a color image encryption method is proposed. The results show that the method can encrypt images safely and effectively and protect the security of image information.

A code summarization generation model fusing multi-structure data

YU Tian-ci, GAO Shang

2024, 46(04): 667-675. doi:

Abstract ( 63 )

PDF (804KB) ( 181 ) 　　

Code summarization can help developers understand the function and implementation of the code. The code summarization generation model can automatically identify the key information in the code and generate relevant summarization to improve the readability and maintainability of the code. Existing code summarization generation models usually only use abstract syntax tree structure information to represent code, resulting in low-quality model-generated summarization. Aiming at this problem, this paper proposes a code summarization generation model that integrates multi-structure data. Firstly, the model adds data flow graph structure information to represent code on the basis of abstract syntax tree. Secondly, in order to capture the global information of the code, the model uses Transformer's encoder to encode the abstract syntax tree sequence. In addition, the model uses the graph neural network to extract features from the data flow graph and provide information such as the computational depen- dencies between variables. Finally, the model uses the cross-modal attention mechanism to fuse the two features of the abstract syntax tree and the data flow and generate corresponding summarization through the Transformer decoder. The experimental results show that, compared with the six mainstream models, the model improves the scores of BLEU,METEOR and ROUGE-L on the Java and Python datasets, and the generated summarization is also very readable.

Fuzzy computation tree logic* model checking based on fuzzy measures

LIU Zi-yuan, MA Zhan-you, LI Xia, GAO Ying-nan, HE Na-na, HUANG Rui-qi

2024, 46(04): 676-683. doi:

Abstract ( 55 )

PDF (552KB) ( 186 ) 　　

A fuzzy computation tree logic* model checking algorithm based on fuzzy measures is proposed for the verification problem of complex systems with fuzziness and uncertainty. Firstly, the syntax and semantics of fuzzy computation tree logic* are introduced based on fuzzy decision process model. Secondly, a fuzzy computation tree logic* model checking algorithm is proposed, which transforms the model checking problem into matrix operation and has the advantages of simplicity of computation and lower complexity. Finally, a medical expert system example is given to illustrate the effectiveness of the model checking algorithm.

A survey of Chinese text classification based on deep learning

GAO Shan, LI Shi-jie, CAI Zhi-ping

2024, 46(04): 684-692. doi:

Abstract ( 245 )

PDF (1058KB) ( 441 ) 　　

In the era of big data, with the continuous popularization of social media, various text data are growing in the network and in life. It is of great significance to analyze and manage text data using text classification technology. Text classification is a basic research field in the field of artificial intelligence natural language processing. Under the given criteria, it classifies text according to content. The application scenarios of text classification are very extensive, such as sentiment analysis, topic classification, relationship classification, etc. Deep learning is a method of representation learning based on data in machine learning, and it shows good classification effect in text data processing. Chinese text and English text have differences in form, sound, and image. Focusing on the uniqueness of Chinese text classification, this paper analyzes and expounds the deep learning methods used for Chinese text classification, and finally sorts out commonly used datasets for Chinese text classification.

An improved snake optimization algorithm based on hybrid strategies and its application

LIANG Xi-ming, SHI Lan-yan, LONG Wen

2024, 46(04): 693-706. doi:

Abstract ( 101 )

PDF (1058KB) ( 201 ) 　　

To solve the problem that the basic snake optimization algorithm easily falls into local optimization, an improved snake optimization algorithm (SSO) based on dimension selection strategy, selection mating strategy, and re-grouping strategy is proposed. The SSO algorithm introduces the dimension selection strategy in the combat or mating stage of the basic snake optimization algorithm. The random probability is used to select the position update mode of each snake individual in different dimensions, so as to avoid the phenomenon of individual position stagnation in the later stage of iteration. The selection mating strategy is introduced in the combat or mating stage, and a part of individuals with smaller fitness values are selected for combat or mating. The remaining individuals use the exploration stage position update formula for position update to improve the exploration ability of the combat or mating stage. The re-grouping strategy is used, and the individuals are randomly disrupted and re-grouped every ten iterations to increase population diversity and improve the optimization ability of the algorithm. Numerical experiments on 30 standard unconstrained optimization problems show that compared with six comparative algorithms such as the basic snake optimization algorithm SO, the SSO algorithm has stronger optimization ability and is more effective for solving high-dimensional optimization problems. The SSO algorithm is used to optimize the initial weights and thresholds of BP neural networks. Experimental results show that the SSO-BP neural network has better accuracy and stability than other comparative neural networks in classifying wines and predicting abalone age.

A dual-view contrastive learning-guided multi-behavior recommendation method

LI Qing-feng, JIN Liu, MA Hui-fang, ZHANG Ruo-yi

2024, 46(04): 707-715. doi:

Abstract ( 59 )

PDF (875KB) ( 169 ) 　　

Multi-behavior recommendation (MBR) typically utilizes various types of user interaction behaviors (such as browsing, adding to cart, and purchasing) to learn user preferences for the target behavior (i.e., purchasing). Due to the impact of sparse supervision signals, existing MBR models often suffer from poor recommendation performance. Recently, contrastive learning has achieved success in mining auxiliary supervision signals from raw data itself. Inspired by this, we propose a dual-view con- trastive learning-guided method to enhance MBR. Firstly, we construct two views that can capture both local and high-order structural information using multi-behavior interaction data. Then, we design two different view encoders to learn user and item embeddings on these complementary views. Finally, we use cross-view collaborative contrastive learning to mutually supervise and learn better embeddings. Experimental results on two real-world datasets demonstrate that our proposed method significantly outperforms baseline methods.

Implicit discourse relation recognition with multi-view contrastive learning

WU Yi-heng, LI Jun-hui, ZHU Mu-hua

2024, 46(04): 716-724. doi:

Abstract ( 60 )

PDF (710KB) ( 157 ) 　　

Previous researches on implicit discourse relationship recognition (IDRR) usually focus on designing effective discourse encoders. Different from theirs, this paper proposes a novel approach which introduces contrastive learning into IDRR so as to obtain representations of discourse units (DUs) with more differentiation. Specifically, a lightweight IDRR classification model is firstly adopted. Then, to better learn representations of DUs, the application of three different contrastive learning methods in IDDR are explored from multiple views, including instance-level, batch-level, and group- level. Finally, three multi-view contrastive learning objectives are combined for better IDRR. Our proposed method only slightly increases training time and introduces small additional parameters. Experimental results on PDTB 2.0 show that our method achieves the state-of-the-art performance.

Efficiency measurement of attribute granulation under the background of three-way concept

ZHANG Xiao-yan, WANG Jia-yi

2024, 46(04): 725-733. doi:

Abstract ( 44 )

PDF (558KB) ( 133 ) 　　

The three-way concept analysis is a combination of three-way decision and formal concept analysis. The greatest progress of this theory compared with formal concept analysis is that it can simultaneously study the information that is “commonly shared” and “not commonly shared” in the formal context. Attribute granu-lation is a theory based on the decomposition of attributes into sub-attributes using a granularity tree and pruning, forming a new set of attributes. However, due to the numerous prunings on the same granularity tree, the key issue to ensure the efficiency of attribute granulation is how to choose the pruning and determine the optimal direction for further operations to achieve optimal granulation results. In this paper, through theoretical derivations, it is proved that there is a close internal relationship between the original three-way concepts and the new three-way concepts obtained from attribute granulation, which can be used as the basis for measuring the efficiency of attribute granulation. Firstly, based on the relationship of attribute granulation levels, the attribute granulation levels are divided into attribute granulation levels with partial order relationships and attribute granulation levels without partial order relationships. Furthermore, the definition of refinement coefficients is given, and the measurement roles of refinement coefficients in the two types of attribute granulation levels are respectively explained, so as to achieve the purpose of measuring the efficiency of different attribute granulation directions.

Date-aware sequential recommendation fusing local information of sequences

CAO Hao-dong, WANG Hai-tao, HE Jian-fen

2024, 46(04): 734-742. doi:

Abstract ( 54 )

PDF (684KB) ( 191 ) 　　

The sequence recommendation algorithm based on self-attention mechanism utilizes users’ interactive sequences to model their dynamic preferences and predict their future behaviors. However, directly inputting the interactive sequences into the self-attention layer will limit the effective utilization of local association information in the sequences. In addition, most of the existing recommendation algorithms use the dot product of the representation of the users’ recent behaviors and the target items to calculate the item scores, which will weaken the impact of previous interactive items on the recommendation results. This paper proposes a date-aware sequential recommendation algorithm that fuses local information of sequences. It uses multiple vertical filters to fuse multiple local association information of each interactive item in the interactive sequence, and uses cross-attention mechanism to capture the relationships between all historical interactive items and target items. It also abandons the traditional position embedding method and replaces it with absolute time embedding based on the date of inter-action occurrence. Experimental results on multiple public datasets show that the algorithm has certain improvement compared with the baseline algorithms in different evaluation indicators.

A neural machine translation method based on language model distillation

SHEN Ying-li, ZHAO Xiao-bing,

2024, 46(04): 743-751. doi:

Abstract ( 63 )

PDF (787KB) ( 174 ) 　　

The lack of large parallel corpora is one of the key issues in low-resource neural machine translation. This paper proposes a neural machine translation method based on language model distillation, which regularizes neural machine translation training by using a monolingual language model. This method introduces prior knowledge contained in the language model to improve translation results. Specifically, we draw on the idea of knowledge distillation, and use the target-side language model (teacher model) trained on rich monolingual data to construct the regularization factor of the low- resource neural machine translation model (student model), allowing the translation model to learn highly generalized prior knowledge from the language model. Unlike traditional monolingual language models that participate in the decoding process, the language model in this method is only used during training and does not participate in the inference stage, so it can effectively improve decoding speed. Experimental results on two low-resource translation datasets of Uyghur-Chinese and Tibetan-Chinese from the 17th national machine translation conference (CCMT2021) show that compared with the current state-of-the-art language model fusion baseline system, BLEU can be improved by 1.42 points (Tibetan-Chinese) to 2.11 points (Chinese-Uyghu).

A microblog rumor detection model based on user authority and multi-feature fusion

XU Li-fen, CAO Zhan-mao, ZHENG Ming-jie, XIAO Bo-jian

2024, 46(04): 752-760. doi:

Abstract ( 94 )

PDF (718KB) ( 228 ) 　　

The widespread dissemination of online rumors and their negative impact on society urgently require efficient rumor detection. Due to the lack of semantic information and strict syntactic structure in the text of the dataset, it is meaningful to combine user characteristics and contextual features to enrich semantic information. In this regard, MRUAMF is proposed. Firstly, four indicators including user information completeness, user activity, user communication span, and user platform authentication index are extracted to construct a quantitative calculation model for user authority. By cascading user authority and its constituent indicators, and using a two-layer fully connected network to fuse features, user characteristics are effectively quantified. Secondly, considering the effectiveness of context in understanding rumors, relevant contextual features are extracted. Finally, the BERT pre-training model is used to extract text features, which are then combined with the Multimodal Adaptation Gate (MAG) to fuse user features, contextual features, and text features. Experiments on the microblog dataset show that compared with the baseline model, the MRUAMF model has better detection performance with an accuracy rate of 0.941.

Current Issue

Author center

Review center

Online journal