Home Latest Articles
Latest Articles
  • Shiyan Luo, Sicong Liu, Bin Guo, Cheng Fang, Minfan Wang, Sai Guo, Zhiwen Yu
    Journal of Computer Research and Development. 2025, 62(12): 3122-3146.

    In recent years, it has become a trend to introduce deep neural network (DNN) into mobile devices. Many applications that facilitate daily life, such as voice assistants and activity recognition, have been integrated into smartphones, wearable devices, and embedded systems. However, it is challenging to deploy CPU-bound DNN on mobile devices with limited resources, such as computing power, storage, and battery. Existing methods, such as manually designed DNN compression techniques and automated on-demand DNN compression techniques, are limited to optimizing model structures. It restricts the upper limit of performance optimization for DNN deployment and makes it difficult to adapt to devices with extremely constrained resources. In addition, these statically pre-designed optimization methods do not consider the resource contention and dynamic demand characteristics of the deployment environment in mobile applications. The inability to adjust strategies in real-time under dynamic environment results in suboptimal accuracy-efficiency performance. To address these challenges, we propose AdaInfer, a runtime-scalable cross-layer optimization method for DNN. AdaInfer adaptively selects the optimal comprehensive deployment strategy for model layers, computational graph layers, and memory layers based on current hardware resource constraints and user performance requirements to optimize multiple performance metrics. It also adjusts the optimal strategy in real-time as the scenario changes. Specifically, we design a scalable graph-computation structure that is model-agnostic and a corresponding cross-layer optimization strategy. These are capable of automatically adjusting to maximize deployment efficiency on heterogeneous devices. Then, we model the runtime adjustment problem of the algorithm-system cross-layer optimization strategy as a dynamic optimization problem and represent the dynamic environment through a set of runtime-varying resource constraints. We also propose an efficient search strategy to enhance the efficiency and quality of local online searches. In evaluations conducted on three types of mobile and edge devices, five models, and four continuously changing mobile scenarios, experimental results show that compared with previous work, AdaInfer reduces memory usage by up to 42.35% and latency by up to 73.89% without significantly affecting accuracy.

  • Hao Wang, Yong Wang, Changlei Feng, Weixin Gai, Peng Wu, Jiang Qian
    Journal of Computer Research and Development. 2025, 62(11): 2651-2662.

    As the breaker of Moore’s Law, Chiplet technology has high expectations in the integrated circuit industry. Chiplet technology can combine multiple small chips with specific functions into a Chiplet integrated chip through high-speed interconnection technology, whose core technology is the Chiplet interconnection technology that can achieve Chiplet combination and expansion. We analyze and discuss the Chiplet interconnection protocol, interconnection architecture, fault-tolerant mechanism, typical interconnection Chiplets, and testability design based on the interconnection Chiplet. Firstly, we provide a detailed comparison and analysis of domestic and foreign Chiplet interconnection protocols, and provide the layers and functions of each protocol. Secondly, we introduce three typical Chiplet interconnection architectures, and analyze the characteristics and advantages of each architecture. Afterwards, the Chiplet fault-tolerant mechanism is introduced, including fault-tolerant encoding of interconnection interfaces, fault-tolerant topology, and fault-tolerant routing. Then, three types of interconnection Chiplet design schemes are presented, including programmable interconnection Chiplets, path programmable interconnection Chiplets, and fully customized interconnection Chiplets. Finally, a testability design testing scheme based on the interconnection Chiplet is introduced. This paper focuses on Chiplet interconnection and aims to help readers have a systematic understanding of Chiplet interconnection technology.

  • Zhichao Wang, Liang Chen, Qianpeng Li, Aoxin Chen, Xin Liu, Wenna Song
    Journal of Computer Research and Development. 2025, 62(11): 2663-2675.

    The rise and development of large-scale neuromorphic platforms require network-on-chip to support efficient data transmission mechanisms. Although many efforts have been made to develop high performance topology architectures and routing schemes, they still suffer from single transmission mode or poor scalability, making them stay on a low efficiency in neuromorphic computing. Inspired by the small-world properties of human brain networks, this brief proposes an efficient region-broadcast (ReB) routing scheme to support unicast, multicast, and broadcast transmission modes. Besides, a synaptic connections indexing method is deployed to accommodate the ReB routing scheme and support this hybrid-mode packet transmission. This method replaces the traditional multicast routing table, effectively improving network scalability and reducing power consumption. Experimental results show that compared with existing work, the ReB routing scheme reduces the peak spike traffic and link load standard deviation by 11.5% and 20.4%, respectively. The ReB routing scheme brings improvements in latency, throughput, and energy under the validation of synthetic traffic, spiking neural network applications and brain cortical networks. Various synthetic traffic patterns are used in the experiments. The datasets used in spiking neural network applications include MNIST, QTDB, Ev-object, and DVS-Gesture. Finally, the proposed ReB router has an excellent bandwidth of 0.24 spike/cycle and only consumes an area of 0.014 mm2.

  • Genggeng Liu, Junqian Huang, Yuhan Zhu, Yantao Yu, Ling Wei, Jiarui Chen, Ning Xu
    Journal of Computer Research and Development. 2025, 62(11): 2676-2686.

    The high-performance global routing results can effectively meet the design specifications and greatly improve the efficiency of the detailed routing phase. As the signal transmitted in the chip increases through the bus, the bus gradually becomes a key factor affecting the performance of the chip. If the bus topology structure is not considered during the global routing stage, coupling phenomena will occur when the bus transmission information is transmitted, resulting in a large timing deviation of signal transmission. Therefore, in order to optimize the consistency of bus topology structure in 2D global routing, an effective multi-strategy bus topology aware global routing algorithm is proposed. First, a topology reconstruction strategy based on congestion is designed to optimize the two-pin net and improve the utilization rate of routing space effectively. Second, a heuristic reroute model is constructed to realize the reroute of multi-signal bit bus. Third, a routing algorithm considering topology structure is added to the rip-up and reroute model to adjust the topology structure of the same bus net group and improve the consistency of bus topology. Finally, an iterative method to adjust the cost of bus topology is designed to further optimize the consistency of bus topology. Experimental results show that the proposed algorithm can effectively optimize the bus topology consistency of 2D routing solution.

  • Wenbin Guo, Xin Wang, Zhao Li, Zirui Chen, Jun Zhao, Feng Feng, Fang Du
    Journal of Computer Research and Development. 2025, 62(11): 2687-2696.

    Knowledge graphs often face the challenge of incompleteness, which can be alleviated by completing missing information through link prediction tasks. However, most knowledge graph completion work overly focuses on extracting embedding features without sufficiently considering the complex semantics contained in the predicted node neighborhood information, global feature information, and directional feature information, making it difficult to accurately predict the missing information. We propose a general representation learning semantic enhancement framework, ASFR, which utilizes an attention mechanism to extract local association information of the knowledge graph and structural features of the knowledge graph, and enhances existing knowledge graph representation learning models by incorporating positional information. By embedding these three types of additional knowledge graph information into the entity vectors of the knowledge graph, the quality of the knowledge graph representation vectors is improved. Comparative experiments are conducted using five different categories of classical methods, and the results indicate that this framework can effectively enhance the predictive capability of models, achieving an improvement of 6.89% on three public datasets.

  • Xibin Jia, Chen Li, Luo Wang, Muchen Zhang, Xiaojian Liu, Yangyang Zhang, Jiakai Wen
    Journal of Computer Research and Development. 2025, 62(11): 2697-2709.

    Multimodal sentiment analysis aims to utilize the multimodal customer comments and other data to identify users’ sentimental tendencies. To realize cross-domain application with the domain bias, commonly used solutions are unsupervised domain adaptation methods. Nevertheless, this type of solutions focuses on the extraction of domain-invariant features, and it neglects the significance of domain-specific features at the target domain. Thus, a meta-optimization based domain-invariant and domain-specific feature disentanglement network is proposed. First, by embedding adapters into the pre-trained large model with fine-tuning fitting, the image-text fused sentiment feature encoder is accordingly constructed. Then, a feature disentanglement module is constructed on the basis of the factorization operation, which utilizes domain adversary and domain classification, together with collaborative independence constraints, respectively, to achieve knowledge-transferable domain-invariant feature embedding while extracting the domain-specific features to enhance the performance of sentiment classification at the target domain. To ensure the consistency of the overall optimization tendency for feature disentanglement and sentiment classification, a meta-learning-based meta-optimization training strategy is put forward to synergistically optimize the sentiment analysis network. Comparative experiments on bidirectional sentiment transfer tasks constructed by MVSA and Yelp datasets demonstrate that compared with other advanced image-text sentiment transfer algorithms, the proposed algorithm achieves superior performance on bidirectional sentiment transfer tasks in terms of three consensus metrics: Precision, Recall and F1 score.

  • Xueying Chen, Chuan Luo, Tianrui Li, Hongmei Chen
    Journal of Computer Research and Development. 2025, 62(11): 2710-2724.

    Feature selection is an effective technique of dimensionality reduction in the field of machine learning. In the era of big data, data security has become an issue of great concern nowadays, and how to perform the feature selection task under the premise of privacy protection is a challenging scientific problem that needs to be solved urgently. Rough hypercuboid is an uncertainty approximation computational model combining rough set theory and hypercuboid learning, which provides an efficient feature selection method for numerical approximate classification problems by introducing supervised information granulation technique and multiple feature evaluation criteria. In this paper, we propose a novel multi-party federated feature selection algorithm under privacy protection, based on the rough hypercuboid model and particle swarm optimization algorithm. Firstly, a centralized (client/server) federated feature selection architecture for multi-party participation is established. Based on the architecture, the rough hypercuboid model and the particle swarm optimization algorithm are used to search the optimal feature subset on the client, and a novel global feature subset evaluation strategy for multiple participants is proposed on the server. Then, the ability of the proposed algorithm to select features in collaboration with multiple participants is improved by designing a particle initialization strategy in a federated environment. Finally, experimental results on the twelve UCI benchmark datasets show that compared with the other six traditional feature selection algorithms, the subset of features selected by the proposed algorithm has higher classification performance on each participant under the premise of satisfying the data privacy protection.

  • Junzuo Zhou, Jiangyan Yi, Jianhua Tao, Yong Ren, Tao Wang
    Journal of Computer Research and Development. 2025, 62(11): 2725-2736.

    At present, end-to-end speech neural codecs, represented by SoundStream, have demonstrated outstanding performance in reconstructed speech quality. However, these methods require extensive convolutional computations, leading to lengthy encoding times. To address this issue, we introduce a neural speech codec method based on Mel spectrogram and squeeze-excitation-weighted quantization. This method aims to maintain high speech perceptual quality while reducing computational costs and increasing operational speed, thereby minimizing latency. Specifically, we utilize Mel spectrogram features as input, capitalize on the temporal compression properties during Mel spectrogram extraction, and combine a lower-layer convolutional encoder to simplify the computation process. Additionally, inspired by squeezed excitation network concepts, we extract excitation weights for each dimension of the output features from the encoder’s final layer. These weights are used as the weighting coefficients for each dimension of the compressed features when calculating codebook distances in the quantizer, thus enabling the learning of correlations between features and enhancing the performance of quantization. Experimental results on the LibriTTS and VCTK datasets indicate that this method significantly enhances the computational speed of the encoder and improves the reconstructed speech quality at lower bit rates (≤3 Kbps). For instance, at a bitrate of 1.5 Kbps, the real-time factor (RTF) of encoding computations can increase by up to 4.6 times. Regarding perceptual quality, at a bitrate of 0.75 Kbps, objective metrics such as short-time objective intelligibility (STOI) and virtual speech quality objective listener (VISQOL) show an average improvement of 8.72% compared with the baseline. Additionally, ablation studies not only demonstrate that the optimization effect of compressed excitation weight methods is inversely correlated with bitrate, but also reveal that, compared with the periodic activation function Snake, the Relu activation function can significantly speed up processing while maintaining comparable speech perceptual quality.

  • Xiaotian Cheng, Weiping Ding, Yu Geng, Jiashuang Huang, Hengrong Ju, Jing Guo
    Journal of Computer Research and Development. 2025, 62(11): 2737-2751.

    Transformer has gradually become the preferred solution for computer vision tasks, which has promoted the development of its interpretability methods. Traditional interpretation methods mostly use the perturbation mask generated by the Transformer encoder’s final layer to generate an interpretable map. However, these methods ignore uncertain information on the mask and the information loss in the upsampling and downsampling processes, which can result in rough and incomplete positioning of the object area. To overcome the mentioned problems, a Transformer explanation method based on sequential three-way and attention fusion (SAF-Explainer) is proposed. SAF-Explainer mainly includes the sequential three-way mask (S3WM) module and attention fusion (AF) module. The S3WM module processes the mask by applying strict threshold conditions to avoid the uncertainty information in the mask from damaging the interpretation results, so as to effectively locate the object position. Subsequently, AF module uses attention matrix aggregation to generate a relationship matrix for cross-layer information interaction, which is used to optimize the detailed information in the interpretation results and generates clear and complete interpretation results. To verify the effectiveness of the proposed SAF-Explainer, comparative experiments are conducted on three natural image datasets and one medical image dataset. The results show that SAF-Explainer has better explainability. This work advances visual explanation techniques by providing more accurate and clinically relevant interpretability for Transformer-based vision systems, particularly in medical diagnostic applications where precise region identification is crucial.

  • Mingshi Yan, Huilin Chen, Zhiyong Cheng, Yahong Han
    Journal of Computer Research and Development. 2025, 62(11): 2752-2766.

    Collaborative filtering-based recommender systems that only rely on single-behavior data often encounter serious sparsity problems in practical applications, resulting in poor performance. Multi-behavior recommendation (MBR) is a method that seeks to learn user preferences, represented as vector embeddings, from auxiliary behavior interaction data. By leveraging these preferences for target behavior recommendations, MBR can mitigate the data sparsity challenge and enhance predictive precision for recommendations. We introduce MB-HGCN, a novel recommendation method designed to exploit multi-behavior data. The method leverages a hierarchical graph convolutional network to learn user and item embeddings from a coarse-grained global level to a fine-grained behavior-specific level. Our method learns global embeddings from a unified homogeneous graph constructed by the interactions of all behaviors, which are then used as initialized embeddings for behavior-specific embedding learning in each behavior graph. Moreover, we also emphasize the distinct of the user and item behavior-specific embeddings and design two simple-yet-effective strategies to aggregate the behavior-specific embeddings for users and items, respectively. Finally, we adopt multi-task learning for optimization. Extensive experimental results on three real-world benchmark datasets show that MB-HGCN can substantially outperform the state-of-the-art methods, achieving a relative improvement of 73.93% and 74.21% for HR@10 and NDCG@10, respectively, on Tmall dataset.