Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2021 Vol.34 Issue.12, Published 2021-12-25 Papers and Reports    Researches and Applications    Surveys and Reviews        Papers and Reports 1069 Learning Paths and Skills Assessment in Formal Context ZHOU Yinfeng, LI Jinjin, FENG Danlu, YANG Taoli In the learning process, learners may learn and master some skills with their knowledge state unchanged. In this situation, learners' skills cannot be assessed accurately due to the unchanged knowledge state. In this paper, a method of formal concept analysis is employed based on the skill function to find learning paths and conduct skills assessment. Firstly, the concepts of subsequent state, effective skill and well-formed skill function are introduced. Secondly, based on the formal context, the conditions that the skill functions satisfy the well-formedness are discussed in two situations. The results of gradual effective learning and effective assessment are obtained under the well-formedness conditions, and the algorithms for obtaining the well-formed skill contexts and the well-formed skill functions and finding learning paths are designed. Finally, the effectiveness of the proposed algorithms is verified on two datasets. The learning paths diagram obtained by the well-formed skill function can not only guide the learners to study effectively but also evaluate whether the learners master the corresponding effective skills according to the change of the learners' knowledge states. 2021 Vol. 34 (12): 1069-1084 [Abstract] ( 830 ) [HTML 1KB] [ PDF 877KB] ( 412 ) 1085 Optimal Transport Based Hierarchical Graph Kernel MA Kai, HUANG Shuo, ZHANG Daoqiang In the existing graph kernels, local attributes of graphs are concerned and local topological features are utilized to compute the similarity measurement of graphs. However, hierarchical structure information of the graph is ignored. To handle this problem, optimal transport based hierarchical graph kernel is proposed. Firstly, each graph is represented as a hierarchical graph structure. During the constructive process of hierarchical graph structure, K-means clustering algorithm is employed to construct new nodes and probabilities of connections between new nodes is regarded as edges of graph at each layer. Then, the optimal transport distance between paired graphs in the special hierarchical structure is calculated using optimal transport with entropic constraints. Finally, the optimal transport distance based hierarchical graph kernel is calculated. The experimental results on six graph datasets show that the classification performance is significantly improved by the proposed graph kernel. 2021 Vol. 34 (12): 1085-1092 [Abstract] ( 372 ) [HTML 1KB] [ PDF 930KB] ( 382 ) 1093 Meta-Path and Hierarchical Attention Based Temporal Heterogeneous Information Network Representation Learning QIN Haiying, ZHAO Zhongying, LI Jianhui, LI Chao Heterogeneous information network representation learning is widely applied in many fields including node classification, link prediction and personalized recommendation. The existing heterogeneous information network representation learning methods mainly focus on static networks but ignore the influence of time on node representations. To address this problem, a meta-path and hierarchical attention based temporal heterogeneous network representation learning method is proposed. The meta-paths are utilized to capture the structural and semantic information in heterogeneous information networks. Through the time decay attention layer, the impact of different meta-path instances at a specific time on the target node is captured. Through the meta-path level attention, the node representation under different meta-paths is fused to obtain the final representation. The experiments on DBLP and IMDB datasets show that the proposed method achieves better results on the tasks of node classification and node clustering. 2021 Vol. 34 (12): 1093-1102 [Abstract] ( 417 ) [HTML 1KB] [ PDF 1021KB] ( 500 ) Surveys and Reviews 1103 3D Object Detection Based on Convolutional Neural Networks: A Survey WANG Yadong, TIAN Yonglin, LI Guoqiang, WANG Kunfeng, LI Dazi Three-dimensional(3D) object detection plays a critical role in the fields of autonomous driving and robotics, since deep learning methods can offer possible solutions for accurate object detection, especially convolutional neural networks. The research progresses of convolutional neural network-based 3D object detection are reviewed comprehensively. Firstly, the practical value, basic procedures and challenges of 3D object detection are summarized. Next, the preliminary knowledge of convolutional neural networks, typical 2D object detection network structures, some widely-used open source datasets and point cloud representations is introduced. Then, progresses on the application of convolutional neural networks in 3D object detection are presented, and the methods are sorted out and analyzed according to different data modalities and method commonalities. Finally, issues in the existing research of 3D object detection are discussed, and future research trends are prospected. 2021 Vol. 34 (12): 1103-1119 [Abstract] ( 1036 ) [HTML 1KB] [ PDF 1041KB] ( 664 ) 1120 Review on Multi-granulation Computing Models and Methods for Decision Analysis PANG Jifang, SONG Peng, LIANG Jiye As the core concept and key technology of granular computing, multi-granulation computing emphasizes multi-view and multi-level understanding and description of real-world problems to obtain more reasonable and satisfactory results. The existing four types of multi-granulation computing models are firstly introduced, including multi-granulation rough set, multi-scale data analysis, sequential three-way decision and hierarchical classification learning, for the further effective fusion of multi-granulation computing and decision analysis and better satisfaction with actual decision-making needs. Then, their main characteristics and development process are expounded. Furthermore, the research status of decision analysis methods based on multi-granulation computing models is summarized from the aspects of attribute reduction, rule extraction, granularity selection, information fusion, group decision-making, multi-attribute group decision-making, classification decision-making and dynamic decision-making. Finally, some challenging research directions of intelligent decision-making in the era of big data are forecasted to promote the continuous development and innovation of multi-granulation intelligent decision-making. 2021 Vol. 34 (12): 1120-1130 [Abstract] ( 652 ) [HTML 1KB] [ PDF 748KB] ( 443 ) Researches and Applications 1131 Local Online Streaming Feature Selection Based on Max-Decision Boundary SUN Shiming, DENG Ansheng The existing online streaming feature selection algorithms usually select the optimal global feature subset, and it is assumed that this subset adapts to all regions of the sample space. However, each region of the sample space is characterized accurately by its own distinct feature subsets. The feature subsets are likely to be different in feature and size. Therefore, an algorithm of local online streaming feature selection based on max-decision boundary is proposed. The local feature selection is introduced. With the full usage of local information, feature measurement standards based on max-decision boundary are designed to separate samples of the same class from samples of different classes as far as possible. Meanwhile, three strategies, maximizing average decision boundary, maximizing decision boundary and minimizing redundancy, are employed to select appropriate features. The class similarity measurement method is applied after the optimal feature subset is selected for the local regions. Experimental results and statistical hypothesis tests on fourteen datasets demonstrate the effectiveness and stability of the proposed algorithm. 2021 Vol. 34 (12): 1131-1142 [Abstract] ( 377 ) [HTML 1KB] [ PDF 560KB] ( 290 ) 1143 Low-Bit Quantization of Neural Network Based on Exponential Moving Average Knowledge Distillation LÜ Junhuan, XU Ke, WANG Dong Now the memory and computational cost restrict the popularization of deep neural network application, whereas neural network quantization is an effective compression method. As the number of quantized bits is lower, the classification accuracy of neural networks becomes poorer in low-bit quantization of neural networks. To solve this problem, a low-bit quantization method of neural networks based on knowledge distillation is proposed. Firstly, a few images are exploited for adaptive initialization to train the quantization step of activation and weight to speed up the convergence of the quantization network. Then, the idea of exponential moving average knowledge distillation is introduced to normalize distillation loss and task loss and guide the training of quantization network. Experiments on ImageNet and CIFAR-10 datasets show that the performance of the proposed method is close to or better than that of the full precision network. 2021 Vol. 34 (12): 1143-1151 [Abstract] ( 552 ) [HTML 1KB] [ PDF 681KB] ( 467 ) 1152 Continual Zero-Shot Learning Algorithm Based on Latent Vectors Alignment ZHONG Xiaorong, HU Xiao, DING Jiayu Continual zero-shot learning aims to accumulate the knowledge of seen classes and utilize the knowledge for unseen classes recognition. However, catastrophic forgetting can easily occur in continual learning. Therefore, a continual zero-shot learning algorithm based on latent vectors alignment is proposed. Based on the cross and distribution aligned variational auto-encoder network, the visual latent vectors of current tasks and learned tasks are aligned to enhance the similarity of latent space of different tasks. Selective retraining is adopted to improve the discrimination ability of the current task model for learned tasks. For different tasks, the independent classifiers are trained with visual-hidden vectors of the seen classes and semantic-hidden vectors of the unseen classes to achieve zero-shot image classification. Extensive experiments on four standard datasets show that the proposed algorithm completes the continual zero-shot recognition task effectively and alleviates the catastrophic forgetting. 2021 Vol. 34 (12): 1152-1159 [Abstract] ( 447 ) [HTML 1KB] [ PDF 709KB] ( 259 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn