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....

2020 Vol.33 Issue.1, Published 2020-01-25 Papers and Reports    Researches and Applications        Papers and Reports 1 Transfer Learning with Joint Inter-class and Inter-domain Distributional Adaptation LI Ping, NI Zhiwei, ZHU Xuhui , SONG Juan Inter-domain information is lost during the process of inter-domain distributional adaptation. Therefore, it is difficult to train an effective classifier in the source domain, and the performance of generalization and tagging in the target domain are affected. Aiming at this problem, an approach, joint inter-class and inter-domain distributional adaptation for transfer learning, is proposed to address this challenge. The proposed method is formulated by learning a projection matrix to map new representations of respective domains into a common subspace. And the distance-measure method of the maximum mean discrepancy is adopted to compute the distance of inter-class and inter-domain distributions. During the optimization procedure, the inter-domain distributional difference is reduced explicitly, and the inter-class distributional difference is enlarged greatly. The capability of knowledge transfer between different domains is improved. Experiments on transfer learning dataset verify the effectiveness of the proposed approach. 2020 Vol. 33 (1): 1-10 [Abstract] ( 710 ) [HTML 1KB] [ PDF 766KB] ( 827 ) 11 Generation of Localized and Visible Adversarial Perturbations ZHOU Xingyu, PAN Zhisong, HU Guyu, DUAN Yexin Deep neural network is susceptible to the disturbance of adversarial attacks. Based on the generative adversarial networks, a novel model of GAN for generating localized and visible adversarial perturbation(G2LVAP) is proposed. Firstly, the attacked classification network is designated as a discriminator, and its parameters are fixed during the training process. The generator model is constructed to generate localized and visible adversarial perturbations by optimizing fooling loss, diversity loss and distance loss. The generated perturbations can be placed anywhere in different input examples to attack the classification network. Finally, a class comparison method is proposed to analyze the effectiveness of localized and visible adversarial perturbations. Experiments on public image classification datasets indicate that G2LVAP produces a satisfactory attack effect. 2020 Vol. 33 (1): 11-20 [Abstract] ( 751 ) [HTML 1KB] [ PDF 3581KB] ( 514 ) 21 Attribute Reductions of Fuzzy-Crisp Concept Lattices Based on Matrix LIN Yidong , LI Jinjin, ZHANG Chengling A matrix representation of fuzzy-crisp formal concepts based on fuzzy formal contexts and a matrix approach of attribute reduction are studied. Firstly, the matrix representations of the extension and intension of fuzzy-crisp concept are developed from the matrix perspective, respectively. The definition and the computing method of attribute granular matrix are formulated subsequently. To find the minimal generation group of fuzzy-crisp concept lattice, matrix judgment theorem of meet-irreducible elements is discussed, and it is utilized to construct the attribute reduction framework preserving the extents of meet-irreducible elements. The significance measure of attribute is proposed by introducing the similarity degree between attribute subsets with aforementioned matrices. And then a heuristic matrix-method of attribute reduction is developed. Finally, numerical experiments verify the effectiveness of the proposed approach. 2020 Vol. 33 (1): 21-31 [Abstract] ( 399 ) [HTML 1KB] [ PDF 617KB] ( 318 ) 32 Distributed Singular Value Decomposition Recommendation Algorithm Based on LU Decomposition and Alternating Least Square LI Lin , WANG Peipei , GU Peng , XIE Qing Aiming at the problems of high time complexity and long running time of the current distributed potential factor recommendation algorithm, a distributed singular value decomposition recommendation algorithm based on LU decomposition and alternating least square(ALS) is proposed. Based on the characteristics of ALS for distributed solution of objective function, a grid-like distributed granularity segmentation strategy is proposed to obtain independent and unrelated feature vectors. When the characteristic matrix is updated, LU decomposition is adopted to solve the inverse matrix to speed up the operation of the algorithm. The experiment on Tencent Weibo dataset in KDD CUP 2012 Track1 indicates that the recommendation speed and efficiency of the proposed algorithm is significantly improved on the premise of ensuring a certain recommendation accuracy. 2020 Vol. 33 (1): 32-40 [Abstract] ( 433 ) [HTML 1KB] [ PDF 751KB] ( 389 ) 41 Chunk-by-Chunk Incomplete Multi-view Clustering Based on Orthogonal Constraints JIANG Jianwei, YIN Jun Existing incomplete multi-view clustering algorithms based on nonnegative matrix factorization(NMF) cannot extract local features accuratly. To solve this problem, an algorithm of chunk-by-chunk incomplete multi-view clustering based on orthogonal constraints (CIMVCO) is proposed. A potential feature matrix of all views is obtained by nonnegative matrix factorization, and orthogonal constraints are added to obtain better local features. For missing samples of each view, smaller weights are given to reduce the impact of missing data. To solve the problem of large scale data clustering, data are processed block-by-block to reduce the memory demand and processing time. Experimental results on Reuters and Digit datasets demonstrate the effectiveness of CIMVCO. 2020 Vol. 33 (1): 41-49 [Abstract] ( 448 ) [HTML 1KB] [ PDF 623KB] ( 356 ) Researches and Applications 50 Cognitive Map Building Method Based on Rat Hippocampus Using CNN YU Naigong, WEI Yaqian, WANG Lin Rat hippocampal formation model fusing visual information has problems of low pose estimation accuracy of closed loop detection and inaccurate map construction. Aiming at the problems, a cognitive map building method based on rat hippocampal formation using convolutional neural network(CNN) is proposed. The improved CNN is utilized to extract visual input features. Location information is obtained by integrating spatial cell computing model and the information is fused to construct cognitive map. Hamming distance is adopted to calculate the similarity between visual information and images in visual library, and thus the familiar scene in the complex dynamic environment is recognized, and the self-positioning and position correction of the robot are completed. Simulation and physical experiments indicate that the proposed method is effective and robust. 2020 Vol. 33 (1): 50-58 [Abstract] ( 601 ) [HTML 1KB] [ PDF 831KB] ( 480 ) 59 Block Target Tracking Based on Occlusion Detection and Multi-block Position Information Fusion CHU Jun, WEI Zhen, MIAO Jun, WANG Lu Target tracking cannot effectively determine when the target is occluded and match the template update. Aiming at this problem, a block target tracking algorithm based on occlusion detection and multi-block position information fusion is proposed. Firstly, the target is divided into four blocks, and the four ones are combined with the target as a whole. Since the occlusion has the characteristics of local start and directivity, the ratio of correlation values between each block is calculated to determine whether and where the target is occluded. The update methods are utilized selectively, depending on whether the target is occluded. Finally, the position of the final target is determined according to each unblocked position information. The experiment indicates that the proposed algorithm can effectively determine whether the target is occluded and improve the tracking effect under occlusion. 2020 Vol. 33 (1): 59-65 [Abstract] ( 617 ) [HTML 1KB] [ PDF 971KB] ( 680 ) 66 Object Tracking with Multi-spatial Resolutions and Adaptive Feature Fusion Based on Correlation Filters TANG Zhangyong, WU Xiaojun, ZHU Xuefeng Correlation filter(CF) based trackers cannot take advantage of the complementary characteristic of deep features and shallow features. To mitigate this problem, an object tracking algorithm with multi-spatial resolutions and adaptive feature fusion based on correlation filter is proposed. Firstly, ResNet-50 is employed to extract deep features and enhance the discrimination and robustness of feature representation during tracking. Additionally, according to the characteristic of different features with different spatial resolutions, image patches in different scales are segmented from video frame as the search area to balance the boundary effect and the number of samples. Finally, an adaptive feature fusion strategy is introduced to fuse the response maps corresponding to two kinds of features with adaptive weights to utilize the complementary characteristic. The experiments on multiple standard datasets verify the effectiveness and robustness of the proposed algorithm. 2020 Vol. 33 (1): 66-74 [Abstract] ( 575 ) [HTML 1KB] [ PDF 2075KB] ( 511 ) 75 Recommendation Algorithm Combining Interrelationship Mining and Collaborative Filtering for Items Cold Start REN Yonggong, SHI Jiaxin, ZHANG Zhipeng Collaborative filtering cannot provide personalized recommendation of new items due to the lack of scoring information for incomplete cold start(ICS) or the absence of scoring information for complete cold start(CCS). To address this problem, a recommendation algorithm combing interrelationship mining and collaborative filtering(CF) is proposed. Firstly, relationship features of items are extracted by using interrelationship mining, and the number of available attributes is expanded according to multiple binary relations between attributes. A neighbor selection approach based on interrelationship mining is proposed to increase the diversity of neighboring items. Finally, CF is integrated to solve CCS and ICS problems and then personalized recommendation of new items is realized. The experiments on two real world datasets indicate that the proposed algorithm solves the new item cold start problems of recommender systems effectively. 2020 Vol. 33 (1): 75-85 [Abstract] ( 478 ) [HTML 1KB] [ PDF 691KB] ( 422 ) 86 Propagation Source Tracing Algorithm Based on Priori Estimation YU Huan, ZHANG Sunxian, LIU Ziang, WANG Zhixiao The deficiencies of most existing spreading source detection methods are low detection rate and large error distance due to the disregard of priori estimation of source node. The infection model, susceptible-infected(SI), is utilized to simulate the propagation process of information in weighed social networks, and a spreading source tracing algorithm based on priori estimation is proposed. Both infected and uninfected nodes of neighborhood are considered, and priori estimations are assigned to the source node according to the number relationship between infected nodes and uninfected nodes. Experiments on artificial and real networks indicate that the proposed algorithm achieves a high detection rate, a small error distance and a high accuracy of real source node ranking. 2020 Vol. 33 (1): 86-92 [Abstract] ( 492 ) [HTML 1KB] [ PDF 697KB] ( 365 )

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