Home Latest Articles
Latest Articles
  • Hongliang Wang, Zhenhua Guo, Qi Mu, Shengcai Lu, Yaming Xu, Quanquan Yu
    Journal of Computer Research and Development. 2026, 63(1): 28-40.

    In recent years, although Chinese domestic field programmable gate array (FPGA) manufacturers have developed rapidly, they still face challenges when deploying FPGA heterogeneous accelerators in data centers. Compared with international manufacturers such as Xilinx (now AMD) and Intel, domestic manufacturers generally lack solutions for high-speed transmission between PCIe devices and hosts, especially in the field of high-performance direct memory access (DMA) controller design, where there are obvious shortcomings. To solve this problem, we design and implement a PCIe-based multi-channel chained DMA controller. By using an independent descriptor controller to manage each channel, sharing data movers, and reducing the consumption of FPGA logic resources, this design improves resource efficiency. The adoption of a chain structure for descriptor management reduces CPU interrupt pressure while meeting the requirements for continuous high-speed transmission between hosts and devices. An innovative architecture for asynchronous internal information pre-processing is developed, enabling data stream processing that significantly improves bandwidth utilization and transmission performance. Testing results show that under PCIe Gen3x8, the DMA bandwidth between the hosts and the domestic FPGA accelerator reaches 6.91 GBps (86% utilization rate), supporting up to 16 channels with channel balancing implementation. This design effectively enables large-scale deployment of domestic FPGA heterogeneous accelerators in data center scenarios.

  • Chenyang Li, Chaohong Ma, Xiaofeng Meng
    Journal of Computer Research and Development. 2026, 63(1): 41-65.

    The booming of artificial intelligence (AI) plays a crucial role in various fields (e.g., healthcare, bioinformatics, and financial services). The main paradigm of AI applications involves building machine learning models to explore insights and patterns within data, which are then utilized for reasoning and decision-making. The effectiveness and efficiency of AI systems hinge on two key factors: the first is the model aspect (model-centric), involving the enhancement of network structure like the transitioning from RNN to LSTM, and the optimization of model hyperparameters; the second is the data aspect (data-centric), such as standardizing data formats, increasing data size, reducing data noises. For a long time, debugging AI systems primarily focuses on optimizing models. However, the rise of the digital era, characterized by activities such as social networking and e-commerce, has produced vast and varied data sets, making model-centric debugging fall short of fulfilling the requirements of modern applications. Consequently, both research and industry communities have redirected their focus from models to data to address this gap. To this end, “data debugging” has emerged. Rather than tuning model, data debugging places emphasis on scrutinizing data, which understands the impact of data errors on downstream tasks at each stage of machine learning pipeline, and then debug the corresponding errors to improve model performance. On the basis of a comprehensive survey of data debugging related work, the research framework for data debugging is first proposed. According to the interaction between data debugging methods and machine learning pipeline, the existing methods are categorized into closed data debugging, immersive data debugging, and hybrid data debugging. Then, we thoroughly provide a detailed overview of related work in this field. Next, we empirically evaluate data debugging methods, and summarize the commonly used datasets and evaluation metrics in this research area. Finally, the challenges and future directions of data debugging are pointed out.

  • Meixia Wang, Huiying Wang, Miaomiao Huang, Yuhai Zhao
    Journal of Computer Research and Development. 2026, 63(1): 66-89.

    Multi-instance multi-label learning (MIML) is an important machine learning framework designed to address complex tasks where each object is represented as a bag of multiple instances and can be associated with multiple labels simultaneously. It has been widely applied in domains such as drug activity prediction and medical image analysis. Different from traditional multi-instance learning and multi-label learning, MIML simultaneously captures the diversity in both the input space (multiple instances) and the output space (multiple labels), enabling more comprehensive modeling and understanding of real-world complex objects. However, this expressiveness also brings significant challenges, including high computational complexity, difficulties in model optimization, and limited generalization capabilities. Therefore, MIML has attracted increasing attention from the research community. Despite this growing interest, a comprehensive survey of MIML is still lacking. We first introduce the formal problem definition and notations related to MIML. Then, from the perspective of data complexity, we categorize MIML problems into three main types: standard MIML, multi-source MIML, and MIML with imprecise labeling. These are further refined into nine subcategories based on solution strategies, instance sources, and label completeness, which are systematically discussed and analyzed. We also summarize commonly used datasets and benchmark comparisons for MIML methods. Finally, we present five representative application scenarios of MIML in real-world tasks, outline four promising future research directions, and conclude the paper.

  • Haiwei Pan, Fengming Yu, Kejia Zhang, Haiyan Lan, Qingyu Meng, Zhe Li
    Journal of Computer Research and Development. 2026, 63(1): 90-122.

    Knowledge distillation, as a key technique in deep learning, achieves model compression and acceleration by transferring knowledge from a large teacher model to a smaller student model. Under the premise of maintaining performance, this technology significantly reduces the requirements of computational resources and storage, and facilitates the deployment of high-performance models on resource-constrained edge devices. Firstly, we provide a systematic review of the recent research in knowledge distillation and categorize it from two perspectives: the type of knowledge and teacher-student model architectures. We comprehensively summarize the distillation methods based on three typical types of knowledge: output feature knowledge, intermediate feature knowledge, and relational feature knowledge, as well as distillation methods based on CNN to CNN architecture, CNN to ViT (vision Transformer) architecture, ViT to CNN architecture, and ViT to ViT architecture. Next, we explore various learning paradigms such as offline distillation, online distillation, self-distillation, data-free distillation, multi-teacher distillation, and assistant distillation. Then, we summarize distillation optimization methods based on the distillation process, knowledge structure, temperature coefficient, and loss functions and analyze improvements in distillation brought by adversarial techniques, automated machine learning, reinforcement learning, and diffusion models, and conclude with the implementation of distillation technology in common applications. Despite significant advancements in knowledge distillation, numerous challenges remain in both practical applications and theoretical research. Finally, we provide an in-depth analysis of these issues and offer insights into future development directions.

  • Zituo Li, Jianbin Sun, Guangzhou Chen, Xinyue Fang, Ruijing Cui, Zhiliang Tian, Zhen Huang, Kewei Yang
    Journal of Computer Research and Development. 2026, 63(1): 123-146.

    In recent years, large language models (LLMs) have made significant strides in the field of natural language processing (NLP), demonstrating impressive capabilities in both language understanding and generation. However, despite these advancements, LLMs still face numerous challenges in practical applications. One such issue that has garnered extensive attention from both the academic and industrial communities is the problem of hallucinations. Effectively detecting hallucinations in large language models is a critical challenge for ensuring their reliable, secure, and trustworthy application in downstream tasks such as text generation. We provide a comprehensive review of methods for detecting hallucinations in large language models. Firstly, we introduce the concept of large language models and clarify the definition and classification of hallucinations. Then we systematically examine the characteristics of LLMs throughout their entire lifecycle, from construction to deployment, and delve into the mechanisms and causes of hallucinations. Secondly, based on practical application requirements and considering factors such as model transparency in different task scenarios, we categorize hallucination detection methods into two types: those for white-box models and those for black-box models. A focused review and in-depth comparison of these methods are provided. Then we analyze and summarize the current mainstream benchmarks for hallucination detection, laying a foundation for future research in this area. Finally, we identify various potential research directions and new challenges in the detection of hallucinations in large language models.

  • Yiming Li, Yaxin Yu, Zhisheng Yu, Yiting Si, Yusong Ye
    Journal of Computer Research and Development. 2026, 63(1): 147-161.

    Multimodal recommendation systems aim to deliver more accurate and personalized recommendations by leveraging diverse data modalities such as text and images. Despite their potential, existing approaches face several critical challenges. 1) Feature distortion. The input embeddings are processed by relatively small pre-trained language models and deep convolutional neural networks, which may lead to inaccurate and incomplete feature representations. 2) Single encoding perspective. The multimodal encoding layers of current models only consider encoding from a single perspective that is either memory-based or expansion-based, which limits representational capacity and leads to information loss. 3) Poor multimodal alignment. Embeddings from different modalities are distributed in distinct and heterogeneous semantic spaces, and need to be effectively mapped into a shared semantic space to achieve accurate alignment. However, existing methods typically rely on simple operations such as direct product of behavioral information, which cannot adequately capture the complex and high-order relationships among modalities. As a result, precise alignment across multiple modalities is difficult to achieve. To address these issues, we propose a novel model named DPRec. This model considers encoding from both memory and expansion perspectives and introduces hypergraphs for multi-level precise cross-modal alignment. The model is evaluated through extensive experiments on three real-world datasets. Experimental results confirm the effectiveness of DPRec, demonstrating its superiority over state-of-the-art approaches in improving recommendation accuracy.

  • Delü Lin, Kun He
    Journal of Computer Research and Development. 2026, 63(1): 162-175.

    With the rapid progress of deep learning technology and the continuous exploration of massive datasets, the self-attention module has been widely applied in various fields such as natural language processing, computer vision, and large language models. Although the self-attention module significantly improves the detection accuracy of deep learning models, its huge computational demand makes it particularly difficult to deploy on computing devices with limited computing power. Integer quantization, as one of the key technologies for deploying models on low-power computing chips, faces the problem of high precision loss caused by the structural characteristics of the self-attention module. To address this issue, a thorough analysis of the integer quantization error in the self-attention module is conducted, and two methods, pseudo-softmax vector quantization and block-wise pseudo-softmax vector quantization, are proposed. These two methods aim to significantly improve inference speed while effectively reducing the error caused by integer quantization by performing special integer quantization on the softmax vectors in the self-attention module. Experimental results show that compared with traditional direct quantization methods, the pseudo-softmax vector quantization method can reduce the quantization accuracy loss by 50%, while the block-wise pseudo-softmax vector quantization method can further reduce the accuracy loss by approximately 90%. These results fully demonstrate the effectiveness of the two quantization methods in reducing precision loss, providing strong support for the efficient deployment of the self-attention module on devices with limited computing power.

  • Jiawen Qin, Pengfeng Huang, Qingyun Sun, Haoyi Zhou, Jianxin Li
    Journal of Computer Research and Development. 2026, 63(1): 176-188.

    Self-supervised learning has emerged as a promising approach in addressing the issues of label dependency and poor generalization performance in traditional graph neural networks (GNNs). This method leverages the inherent structure and properties of the data to generate supervisory information without relying on extensive labeled datasets. However, most existing self-supervised learning methods hold the assumption that the graphs are homophilic and consequently fail to generalize well to graphs with high heterophily, where connected nodes may have different class labels and dissimilar features. In this paper, we study the problem by developing an asymmetric self-supervised learning on non-homophilic graphs framework (MIRROR), which is easy to implement and does not rely on random graph augmentations and homophilic assumptions. Inspired by the designs of existing heterophilic graph neural networks, MIRROR learns the node representations by capturing one-hop local neighborhood information and informative distant neighbors. Such two properties are obtained through carefully designed pretext tasks that are optimized based on predicted neighborhood context and estimated high-order mutual information. Extensive experiments on various types of graph benchmarks demonstrate that our proposed framework can achieve better performance compared with competitive baselines on non-homophilic graphs. The superiority in multiple downstream tasks also demonstrates the generalization capability of MIRROR.

  • Chaofan Li, Songcan Chen
    Journal of Computer Research and Development. 2026, 63(1): 189-198.

    Multi-channel electroencephalography (EEG), as a non-invasive technique, records brain electrical activity through multiple electrodes placed on the scalp, aiding in understanding individual psychological states and assisting in the diagnosis of various diseases. Given the high cost and technical challenges of annotating large-scale EEG data, self-supervised learning (SSL), as a label-free learning paradigm, has garnered widespread attention in the EEG domain. SSL leverages the intrinsic structure of data to learn representations, thereby effectively enhancing the model’s generalization ability. Despite significant advancements in current research, the field still faces the following challenges: First, multi-channel EEG data contain complex spatiotemporal correlations, yet many existing methods are confined to modeling either temporal or spatial dimensions in isolation, rather than integrating both comprehensively, thus limiting the ability to profoundly understand the intrinsic complexities of EEG signals. Second, many current approaches have not effectively integrated segment-level and instance-level information. The former contributes to improving the model’s generalization ability, while the latter aids the model in better adapting to downstream (classification) tasks. To address these challenges, we propose and implement a self-supervised pre-training framework that combines contrastive and reconstruction strategies. Specifically, by introducing a channel masking strategy based on temporal mask reconstruction, our framework effectively captures the spatiotemporal relationships of EEG data. By capturing inter-segment relationships in a fine-grained manner, the framework enhances both the model’s performance and its generalization ability. Meanwhile, integrating instance-level contrastive learning with masked reconstruction tasks helps the model learn representations with instance discriminability and local perceptual ability, thus facilitating better adaptation to downstream tasks. Additionally, the introduction of the self-paced learning mechanism further enhances the model’s generalization ability. Finally, experimental results across multiple EEG tasks demonstrate the effectiveness of the proposed approach.

  • Shuangcheng Wang, Songhao Zhu
    Journal of Computer Research and Development. 2026, 63(1): 199-213.

    Few-shot learning is the mainstream method for few-shot object detection, but it has serious shortcomings. 1) The extreme lack of new class samples leads to biased distribution of new class features; 2) Due to the robustness assumption during the fine-tuning process are not necessarily applicable to new class samples, the feature extraction network is unable to extract unbiased new class sample features. To deal with the above two issues, three-stage fine-tuning few-shot object detection method based on cross-module knowledge distillation is proposed. Firstly, the feature distribution calibration strategy is designed to calibrate the feature distribution of new class during the two-step fine-tuning process. Secondly, the proposed first bias reduction strategy, effectively alleviates the bias estimation problem of weight parameters in the linear probing (the first stage of the fine-tuning process), and the proposed inverse first bias reduction effectively alleviates the over-fitting problem of feature extraction network during overall fine-tuning (the second stage of the fine-tuning process). Finally, the proposed cross-module knowledge distillation strategy is utilized to guide the shallow modules of the model to learn deep features to capture more discriminative new class features. A large number of experimental results show that the proposed three-stage fine-tuning few-shot object detection method effectively improves the accuracy and robustness of few-shot object detection.