Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1202131
R. Waivio
In this paper we investigate computational properties of a new N-layer heteroassociative memory model with respect to information encoding. We describe a technique for encoding a set of m/spl times/n matrix patterns where entering one column (row) of a pattern allows the remaining columns (rows) to be recurrently reconstructed. Following are some of the main contributions of this paper: - We show how to transform any given set of patterns to a standard form using a simple procedure. Then we demonstrate that after a competitive initialization among all layers our multilayer network converges in one step to fixed points which are one of the given patterns in its standard form. Due to an increase in the domain of attraction, our architecture becomes more powerful than the previous models. - We analyze the optimal number of layers, as well as their dimensions, based on the cardinality of maximal linearly independent subspaces of the input patterns. - We prove that our proposed model is stable under mild technical assumptions using the discrete Lyapunov energy function.
{"title":"On discrete N-layer heteroassociative memory models","authors":"R. Waivio","doi":"10.1109/ICONIP.2002.1202131","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1202131","url":null,"abstract":"In this paper we investigate computational properties of a new N-layer heteroassociative memory model with respect to information encoding. We describe a technique for encoding a set of m/spl times/n matrix patterns where entering one column (row) of a pattern allows the remaining columns (rows) to be recurrently reconstructed. Following are some of the main contributions of this paper: - We show how to transform any given set of patterns to a standard form using a simple procedure. Then we demonstrate that after a competitive initialization among all layers our multilayer network converges in one step to fixed points which are one of the given patterns in its standard form. Due to an increase in the domain of attraction, our architecture becomes more powerful than the previous models. - We analyze the optimal number of layers, as well as their dimensions, based on the cardinality of maximal linearly independent subspaces of the input patterns. - We prove that our proposed model is stable under mild technical assumptions using the discrete Lyapunov energy function.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125405296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198210
S. Hasegawa, T. Kurita
This paper introduces a method for face and non-face classification. The method is based on the combined use of the multinomial logit model (MLM) and "kernel feature compound vectors". The NMM is one of the neural network models for multi-class pattern classification, and is supposed to be equal or better in classification performance than linear classification methods. The "Kernel Feature Compound Vectors" are compound feature vectors of geometric image features and Kernel features. Evaluation and comparison experiments were conducted by using face and non-ace images (Face training 100, cross-validation 300, test 325, Non-face training 200, cross-validation 1000, test 1000) gathered from the available face databases and others. The experimental result obtained by the proposed method was the best compared with the results by the Support Vector Machines (SVM) and the Kernel Fisher Discriminant Analysis (KFDA).
{"title":"Face and non-face classification by multinomial logit model and kernel feature compound vectors","authors":"S. Hasegawa, T. Kurita","doi":"10.1109/ICONIP.2002.1198210","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198210","url":null,"abstract":"This paper introduces a method for face and non-face classification. The method is based on the combined use of the multinomial logit model (MLM) and \"kernel feature compound vectors\". The NMM is one of the neural network models for multi-class pattern classification, and is supposed to be equal or better in classification performance than linear classification methods. The \"Kernel Feature Compound Vectors\" are compound feature vectors of geometric image features and Kernel features. Evaluation and comparison experiments were conducted by using face and non-ace images (Face training 100, cross-validation 300, test 325, Non-face training 200, cross-validation 1000, test 1000) gathered from the available face databases and others. The experimental result obtained by the proposed method was the best compared with the results by the Support Vector Machines (SVM) and the Kernel Fisher Discriminant Analysis (KFDA).","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"36 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125426619","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198172
H. Kim, T. Lee, D. Lim, D. Jung
We describe the application of a genetic algorithm (GA) to the problem of parameter optimization for an adaptive finite impulse response (FIR) filter combining genetic algorithm (GA) and least mean square (LMS) algorithm. For system identification problem, LMS algorithm computes the filter coefficients and GA search the optimal step-size adaptively. Because step-size influences on the stability and performance, so it is necessary to apply method that can control it. The simulation results of the GA were compared to the traditional LMS algorithm. We obtained that genetic algorithm was clearly superior (in accuracy) in most cases.
{"title":"The hybrid method for determining an adaptive step size of the unknown system identification using genetic algorithm and LMS algorithm","authors":"H. Kim, T. Lee, D. Lim, D. Jung","doi":"10.1109/ICONIP.2002.1198172","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198172","url":null,"abstract":"We describe the application of a genetic algorithm (GA) to the problem of parameter optimization for an adaptive finite impulse response (FIR) filter combining genetic algorithm (GA) and least mean square (LMS) algorithm. For system identification problem, LMS algorithm computes the filter coefficients and GA search the optimal step-size adaptively. Because step-size influences on the stability and performance, so it is necessary to apply method that can control it. The simulation results of the GA were compared to the traditional LMS algorithm. We obtained that genetic algorithm was clearly superior (in accuracy) in most cases.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"69 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126226365","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198218
Yen-Jen Oyang, Shien-Ching Hwang
This paper proposes a novel learning algorithm for constructing function approximators with radial basis function (RBF) networks. In comparison with the existing learning algorithms, the proposed algorithm features lower time complexity for constructing the RBF network and is able to deliver the same level of accuracy. The time taken by the proposed algorithm to construct the RBF network is in the order of O(|S|), where S is the set of training samples. As far as the time complexity for predicting the function values of input vectors is concerned, the RBF network constructed with the proposed learning algorithm can complete the task in O(|T|), where T is the set of input vectors. Another important feature of the proposed learning algorithm is that the space complexity of the RBF network constructed is O(m|S|), where m is the dimension of the vector space in which the target function is defined.
{"title":"An efficient learning algorithm for function approximation with radial basis function networks","authors":"Yen-Jen Oyang, Shien-Ching Hwang","doi":"10.1109/ICONIP.2002.1198218","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198218","url":null,"abstract":"This paper proposes a novel learning algorithm for constructing function approximators with radial basis function (RBF) networks. In comparison with the existing learning algorithms, the proposed algorithm features lower time complexity for constructing the RBF network and is able to deliver the same level of accuracy. The time taken by the proposed algorithm to construct the RBF network is in the order of O(|S|), where S is the set of training samples. As far as the time complexity for predicting the function values of input vectors is concerned, the RBF network constructed with the proposed learning algorithm can complete the task in O(|T|), where T is the set of input vectors. Another important feature of the proposed learning algorithm is that the space complexity of the RBF network constructed is O(m|S|), where m is the dimension of the vector space in which the target function is defined.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"56 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126226751","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1199010
J. Teo, H. Abbass
This paper investigates the use of a multi-objective approach for evolving artificial neural networks that act as controllers for the legged locomotion of a 3-dimensional, artificial quadruped creature simulated in a physics-based environment. The Pareto-frontier Differential Evolution (PDE) algorithm is used to generate a Pareto optimal set of artificial neural networks that optimizes the conflicting objectives of maximizing locomotion behavior and minimizing neural network complexity. Here we provide an insight into how the controller generates the emergent walking behavior in the creature by analyzing the evolved artificial neural networks in operation. A comparison between Pareto optimal controllers showed that ANNs with varying numbers of hidden units resulted in noticeably different locomotion behaviors. We also found that a much higher level of sensory-motor coordination was present in the best evolved controller.
{"title":"Coordination and synchronization of locomotion in a virtual robot","authors":"J. Teo, H. Abbass","doi":"10.1109/ICONIP.2002.1199010","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1199010","url":null,"abstract":"This paper investigates the use of a multi-objective approach for evolving artificial neural networks that act as controllers for the legged locomotion of a 3-dimensional, artificial quadruped creature simulated in a physics-based environment. The Pareto-frontier Differential Evolution (PDE) algorithm is used to generate a Pareto optimal set of artificial neural networks that optimizes the conflicting objectives of maximizing locomotion behavior and minimizing neural network complexity. Here we provide an insight into how the controller generates the emergent walking behavior in the creature by analyzing the evolved artificial neural networks in operation. A comparison between Pareto optimal controllers showed that ANNs with varying numbers of hidden units resulted in noticeably different locomotion behaviors. We also found that a much higher level of sensory-motor coordination was present in the best evolved controller.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"5 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126101998","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198997
L. Tian, Zongyuan Mao
This paper deals with the tracking control problem of a manipulator system with unknown and changing dynamics. In this study, a fuzzy logic controller (FLC) in the feedback configuration is proposed, and an efficient dynamic recurrent neural network (DRNN) in the feedforward configuration is developed. The DRNN, which possesses the ability of approaching arbitrary nonlinear function, is utilized to approximate the inverse dynamics of the robotic manipulator system. Based on the outputs of the FLC, parameter updating equations are derived for the adaptive DRNN model. The analysis of the stability of the system is also carried out. Finally, comparisons between fuzzy control and the proposed controller are carried out. The results demonstrate remarkable performance of the proposed controller for the two-link flexible manipulator system.
{"title":"Fuzzy neuro controller for a two-link rigid-flexible manipulator system","authors":"L. Tian, Zongyuan Mao","doi":"10.1109/ICONIP.2002.1198997","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198997","url":null,"abstract":"This paper deals with the tracking control problem of a manipulator system with unknown and changing dynamics. In this study, a fuzzy logic controller (FLC) in the feedback configuration is proposed, and an efficient dynamic recurrent neural network (DRNN) in the feedforward configuration is developed. The DRNN, which possesses the ability of approaching arbitrary nonlinear function, is utilized to approximate the inverse dynamics of the robotic manipulator system. Based on the outputs of the FLC, parameter updating equations are derived for the adaptive DRNN model. The analysis of the stability of the system is also carried out. Finally, comparisons between fuzzy control and the proposed controller are carried out. The results demonstrate remarkable performance of the proposed controller for the two-link flexible manipulator system.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"42 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116204543","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1202855
Wan Zhang, Irwin King
Recently, support vector machine (SVM) has become a very dynamic and popular topic in the neural network community for its abilities to perform classification, estimation, and regression. One of the major tasks in the SVM algorithm is to locate the points, or rather support vectors, based on which we construct the discriminant boundary in classification task. In the process of studying the methods for finding the decision boundary, we conceive a method, /spl beta/-skeleton algorithm, which reduces the size of the training set for SVM. We describe their theoretical connections and practical implementation implications. In this paper, we also survey four different methods for classification: the SVM method, k-nearest neighbor method, /spl beta/-skeleton algorithm used in the above two methods. Compared with the methods without using /spl beta/-skeleton algorithm, prediction with the edited set obtained from /spl beta/-skeleton algorithm as the training set, does not lose the accuracy too much but reduces the real running time.
{"title":"Locating support vectors via /spl beta/-skeleton technique","authors":"Wan Zhang, Irwin King","doi":"10.1109/ICONIP.2002.1202855","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1202855","url":null,"abstract":"Recently, support vector machine (SVM) has become a very dynamic and popular topic in the neural network community for its abilities to perform classification, estimation, and regression. One of the major tasks in the SVM algorithm is to locate the points, or rather support vectors, based on which we construct the discriminant boundary in classification task. In the process of studying the methods for finding the decision boundary, we conceive a method, /spl beta/-skeleton algorithm, which reduces the size of the training set for SVM. We describe their theoretical connections and practical implementation implications. In this paper, we also survey four different methods for classification: the SVM method, k-nearest neighbor method, /spl beta/-skeleton algorithm used in the above two methods. Compared with the methods without using /spl beta/-skeleton algorithm, prediction with the edited set obtained from /spl beta/-skeleton algorithm as the training set, does not lose the accuracy too much but reduces the real running time.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"120 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121169484","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198220
Tao Wu, Hangen He, D. Hu
How to select a kernel function for the given data is an open problem in the research of support vector machine (SVM). There is a question puzzling many people: suppose the training data are separated nonlinearly in the input space, how do we know that the chosen kernel function can make the training data to be separated linearly in the feature space? A simple method is presented to decide if a selected kernel function can separate the given data linearly or not in the feature space.
{"title":"On the separability of kernel functions","authors":"Tao Wu, Hangen He, D. Hu","doi":"10.1109/ICONIP.2002.1198220","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198220","url":null,"abstract":"How to select a kernel function for the given data is an open problem in the research of support vector machine (SVM). There is a question puzzling many people: suppose the training data are separated nonlinearly in the input space, how do we know that the chosen kernel function can make the training data to be separated linearly in the feature space? A simple method is presented to decide if a selected kernel function can separate the given data linearly or not in the feature space.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116737542","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1198966
L. Litinskii, B. Kryzhanovsky, A. Fonarev
In this paper we develop a formalism allowing us to describe operating of a network based on the parametrical four-wave mixing process that is well-known in nonlinear optics. The recognition power of a network using parametric neurons operating with q different frequencies is considered. It is shown that the storage capacity of such a network is higher compared with the Potts-glass models.
{"title":"Optical neural network based on the parametrical four-wave mixing process","authors":"L. Litinskii, B. Kryzhanovsky, A. Fonarev","doi":"10.1109/ICONIP.2002.1198966","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1198966","url":null,"abstract":"In this paper we develop a formalism allowing us to describe operating of a network based on the parametrical four-wave mixing process that is well-known in nonlinear optics. The recognition power of a network using parametric neurons operating with q different frequencies is considered. It is shown that the storage capacity of such a network is higher compared with the Potts-glass models.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"14 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121376468","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2002-11-18DOI: 10.1109/ICONIP.2002.1202186
T. Nishikawa, S. Abe
According to the CARVE algorithm, any pattern classification problem can be synthesized in three layers without misclassification. In this paper, we propose to train multilayer neural network classifiers based on the CARVE algorithm. In hidden layer training, we find a hyperplane that separates a set of data belonging to one class from the remaining data. Then, we remove the separated data from the training data, and repeat this procedure until only the data belonging to one class remain. In determining the hyperplane, we maximize margins heuristically so that data of one class are on one side of the hyperplane. In output layer training, we determine the hyperplane by a quadratic optimization technique. The performance of this new algorithm is evaluated by some benchmark data sets.
{"title":"Maximizing margins of multilayer neural networks","authors":"T. Nishikawa, S. Abe","doi":"10.1109/ICONIP.2002.1202186","DOIUrl":"https://doi.org/10.1109/ICONIP.2002.1202186","url":null,"abstract":"According to the CARVE algorithm, any pattern classification problem can be synthesized in three layers without misclassification. In this paper, we propose to train multilayer neural network classifiers based on the CARVE algorithm. In hidden layer training, we find a hyperplane that separates a set of data belonging to one class from the remaining data. Then, we remove the separated data from the training data, and repeat this procedure until only the data belonging to one class remain. In determining the hyperplane, we maximize margins heuristically so that data of one class are on one side of the hyperplane. In output layer training, we determine the hyperplane by a quadratic optimization technique. The performance of this new algorithm is evaluated by some benchmark data sets.","PeriodicalId":146553,"journal":{"name":"Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02.","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123890696","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：