Pub Date : 1991-11-18DOI: 10.1109/IJCNN.1991.170439
T. W. Scutt, R. Damper
It is noted that almost all attempts to model neural and brain function have fallen into one of two categories: artificial neural networks using (ideally) large numbers of simple but densely interconnected processing elements, or detailed physiological models of single neurons. The authors report on their progress in formulating a computational model which functions at a level between these two extremes. Individual neurons are considered at the level of membrane potential; this allows outputs from the model to be compared directly with physiological data obtained in intracellular recording. An object-oriented programming language has been used to produce a model where each object equates to a neuron. The benefits of using an object-oriented language are two-fold. The program has been tested by modeling the learning and behavior of the gill-withdrawal reflex in Aplysia. The use of a parameter-based system has made it possible to specify appropriate characteristics for the particular neurons participating in this reflex and to simulate some of the subcircuits involved.<>
{"title":"Computational modelling of learning and behaviour in small neuronal systems","authors":"T. W. Scutt, R. Damper","doi":"10.1109/IJCNN.1991.170439","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170439","url":null,"abstract":"It is noted that almost all attempts to model neural and brain function have fallen into one of two categories: artificial neural networks using (ideally) large numbers of simple but densely interconnected processing elements, or detailed physiological models of single neurons. The authors report on their progress in formulating a computational model which functions at a level between these two extremes. Individual neurons are considered at the level of membrane potential; this allows outputs from the model to be compared directly with physiological data obtained in intracellular recording. An object-oriented programming language has been used to produce a model where each object equates to a neuron. The benefits of using an object-oriented language are two-fold. The program has been tested by modeling the learning and behavior of the gill-withdrawal reflex in Aplysia. The use of a parameter-based system has made it possible to specify appropriate characteristics for the particular neurons participating in this reflex and to simulate some of the subcircuits involved.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115136174","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170574
S. Srirengan, C. Looi
The authors describe the results of initial efforts in applying backpropagation to the prediction of future values of four time series, namely, the sunspot series, a monthly department store sales time series, and two financial index time series. They describe various ways of customizing the backpropagation network for prediction and discuss some experimental results. They also propose a modified learning rule based on optimizing correct predictions of upward and downward trends in a time series.<>
{"title":"On using backpropagation for prediction: an empirical study","authors":"S. Srirengan, C. Looi","doi":"10.1109/IJCNN.1991.170574","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170574","url":null,"abstract":"The authors describe the results of initial efforts in applying backpropagation to the prediction of future values of four time series, namely, the sunspot series, a monthly department store sales time series, and two financial index time series. They describe various ways of customizing the backpropagation network for prediction and discuss some experimental results. They also propose a modified learning rule based on optimizing correct predictions of upward and downward trends in a time series.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125155761","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170666
J. Hao, S. Tan, J. Vandewalle
The authors present a novel method for constructing the weight matrix for the Hopfield associative memory. The most important feature of this method is the explicit introduction of the size of the attraction basin to be a main design parameter, and the weight matrix is obtained as a result of optimizing this parameter. Another feature is that all the connection weights can only assume three different values, -1, +1, and 0, which facilitates the VLSI implementation of the weights. Compared to the widely used Hebbian rule, the method can guarantee all the given patterns to be stored at least as fixed points, regardless of the internal structure of the patterns. The proposed design method is illustrated by a few examples.<>
{"title":"A new approach to the design of Hopfield associative memory","authors":"J. Hao, S. Tan, J. Vandewalle","doi":"10.1109/IJCNN.1991.170666","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170666","url":null,"abstract":"The authors present a novel method for constructing the weight matrix for the Hopfield associative memory. The most important feature of this method is the explicit introduction of the size of the attraction basin to be a main design parameter, and the weight matrix is obtained as a result of optimizing this parameter. Another feature is that all the connection weights can only assume three different values, -1, +1, and 0, which facilitates the VLSI implementation of the weights. Compared to the widely used Hebbian rule, the method can guarantee all the given patterns to be stored at least as fixed points, regardless of the internal structure of the patterns. The proposed design method is illustrated by a few examples.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126183437","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170509
K. Cheung, C. Leung
The authors present a novel architecture, known as the rotational quadratic function neuron (RQFN), to implement the quadratic function neuron (QFN). Although with some loss in the degree of freedom in the boundary formation, RQFN possesses some attributes which are unique when compared to QFN. In particular, the architecture of RQFN is modular, which facilitates VLSI implementation. Moreover, by replacing QFN by RQFN in a multilayer perceptron (MP), the fan-in and the interconnection volume are reduced to that of MP utilizing linear neurons. In terms of learning, RQFN also offers varieties such as the separate learning paradigm and the constrained learning paradigm. Single-layer MP utilizing RQFNs have been demonstrated to form more desirable boundaries than the normal MP. This is essential in the scenario where either the closure of the boundary or boundaries of higher orders are required.<>
{"title":"Rotational quadratic function neural networks","authors":"K. Cheung, C. Leung","doi":"10.1109/IJCNN.1991.170509","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170509","url":null,"abstract":"The authors present a novel architecture, known as the rotational quadratic function neuron (RQFN), to implement the quadratic function neuron (QFN). Although with some loss in the degree of freedom in the boundary formation, RQFN possesses some attributes which are unique when compared to QFN. In particular, the architecture of RQFN is modular, which facilitates VLSI implementation. Moreover, by replacing QFN by RQFN in a multilayer perceptron (MP), the fan-in and the interconnection volume are reduced to that of MP utilizing linear neurons. In terms of learning, RQFN also offers varieties such as the separate learning paradigm and the constrained learning paradigm. Single-layer MP utilizing RQFNs have been demonstrated to form more desirable boundaries than the normal MP. This is essential in the scenario where either the closure of the boundary or boundaries of higher orders are required.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123419141","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170515
K.C. Lee, N. Funabiki, Y.B. Cho, Yoshiyasu Takefuji
A novel computational model for large-scale maximum clique problems is proposed and tested. The maximum clique problem is first formulated as an unconstrained quadratic zero-one programming and it is solved by minimizing the weight summation over the same partition in a newly constructed graph. The proposed maximum neural network has the following advantages: (1) coefficient-parameter tuning in the motion equation is not required in the maximum neural network while the conventional neural networks suffer from it; (2) the equilibrium state of the maximum neural network is clearly defined in order to terminate the algorithm, while the existing neural networks do not have the clear definition; and (3) the maximum neural network always allows the state of the system to converge to the feasible solution, while the existing neural networks cannot guarantee it. The proposed parallel algorithm for large-size problems outperforms the best known algorithms in terms of computation time with much the same solution quality where the conventional branch-and-bound method cannot be used due to the exponentially increasing computation time.<>
{"title":"A parallel neural network computing for the maximum clique problem","authors":"K.C. Lee, N. Funabiki, Y.B. Cho, Yoshiyasu Takefuji","doi":"10.1109/IJCNN.1991.170515","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170515","url":null,"abstract":"A novel computational model for large-scale maximum clique problems is proposed and tested. The maximum clique problem is first formulated as an unconstrained quadratic zero-one programming and it is solved by minimizing the weight summation over the same partition in a newly constructed graph. The proposed maximum neural network has the following advantages: (1) coefficient-parameter tuning in the motion equation is not required in the maximum neural network while the conventional neural networks suffer from it; (2) the equilibrium state of the maximum neural network is clearly defined in order to terminate the algorithm, while the existing neural networks do not have the clear definition; and (3) the maximum neural network always allows the state of the system to converge to the feasible solution, while the existing neural networks cannot guarantee it. The proposed parallel algorithm for large-size problems outperforms the best known algorithms in terms of computation time with much the same solution quality where the conventional branch-and-bound method cannot be used due to the exponentially increasing computation time.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125279669","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170335
Y.-P. Huang, D. Gustafson
The performance of fault tolerant associative memories is investigated. Instead of presenting the results by simulation, the authors mathematically show that the one-step retrieval probability in most cases decreases with the increase in error ratio, number of error bits, and number of stored patterns. For the case of faulty resistance, however, the performance will surpass the nonerror situation under the positive weight change. This is not only true in the Hopfield interconnection topology but is also true in the exponential correlation case.<>
{"title":"Fault tolerant analysis of associative memories","authors":"Y.-P. Huang, D. Gustafson","doi":"10.1109/IJCNN.1991.170335","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170335","url":null,"abstract":"The performance of fault tolerant associative memories is investigated. Instead of presenting the results by simulation, the authors mathematically show that the one-step retrieval probability in most cases decreases with the increase in error ratio, number of error bits, and number of stored patterns. For the case of faulty resistance, however, the performance will surpass the nonerror situation under the positive weight change. This is not only true in the Hopfield interconnection topology but is also true in the exponential correlation case.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115062646","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170764
A. Hartstein
The author investigates the effect of the dimensionality of the interconnections in a Hopfield-type network on the storage capacity of the network. The analysis is performed for 1D, 2D, 3D and 4D interconnection geometries. The capacity was found to be independent of the dimensionality of the interconnections and to depend only on the total number of interconnections available in a given network. In addition, no evidence of any instabilities was observed, in contrast to physical systems of reduced dimensionality.<>
{"title":"The effect of the dimensionality of interconnections on the storage capacity of a threshold controlled neural network","authors":"A. Hartstein","doi":"10.1109/IJCNN.1991.170764","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170764","url":null,"abstract":"The author investigates the effect of the dimensionality of the interconnections in a Hopfield-type network on the storage capacity of the network. The analysis is performed for 1D, 2D, 3D and 4D interconnection geometries. The capacity was found to be independent of the dimensionality of the interconnections and to depend only on the total number of interconnections available in a given network. In addition, no evidence of any instabilities was observed, in contrast to physical systems of reduced dimensionality.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122987483","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170277
C. Chinrungrueng, C. Séquin
A modified k-means competitive learning algorithm that can perform efficiently in situations where the input statistics are changing, such as in nonstationary environments, is presented. This modified algorithm is characterized by the membership indicator that attempts to balance the variations of all clusters and by the learning rate that is dynamically adjusted based on the estimated deviation of the current partition from an optimal one. Simulations comparing this new algorithm with other k-means competitive learning algorithms on stationary and nonstationary problems are presented.<>
{"title":"K-means competitive learning for non-stationary environments","authors":"C. Chinrungrueng, C. Séquin","doi":"10.1109/IJCNN.1991.170277","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170277","url":null,"abstract":"A modified k-means competitive learning algorithm that can perform efficiently in situations where the input statistics are changing, such as in nonstationary environments, is presented. This modified algorithm is characterized by the membership indicator that attempts to balance the variations of all clusters and by the learning rate that is dynamically adjusted based on the estimated deviation of the current partition from an optimal one. Simulations comparing this new algorithm with other k-means competitive learning algorithms on stationary and nonstationary problems are presented.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123012937","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170730
P. Koiran
The author studies the approximation of continuous mappings and dichotomies by one-hidden-layer networks, from a computational point of view. The approach is based on a new approximation method, specially designed for constructing small networks. Upper bounds are given on the size of these networks. These results are specialized to the case of transitional invariant networks, i.e., networks whose outputs are unchanged when their inputs are submitted to a translation.<>
{"title":"Approximations of mappings and application to translational invariant networks","authors":"P. Koiran","doi":"10.1109/IJCNN.1991.170730","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170730","url":null,"abstract":"The author studies the approximation of continuous mappings and dichotomies by one-hidden-layer networks, from a computational point of view. The approach is based on a new approximation method, specially designed for constructing small networks. Upper bounds are given on the size of these networks. These results are specialized to the case of transitional invariant networks, i.e., networks whose outputs are unchanged when their inputs are submitted to a translation.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122108639","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 : 1991-11-18DOI: 10.1109/IJCNN.1991.170707
N. Matsui, E. Bamba
An approach to a macroscopic description of a cluster-formation algorithm by neural activities in a random neural network is considered. The activity interaction between clusters of neurons and the network entropy through the medium of the activity parameter x(p) for the input pattern p, are introduced as a system energy. By using the neural state transition rule similar to that in the Boltzmann network and some simple stochastic assumptions, cluster-formation of neurons was simulated. The relations between cluster sizes, or the simulated activity, and the setting activity parameter are shown. The validity of this macroscopic description is also discussed.<>
{"title":"Neural activities and cluster-formation in a random neural network","authors":"N. Matsui, E. Bamba","doi":"10.1109/IJCNN.1991.170707","DOIUrl":"https://doi.org/10.1109/IJCNN.1991.170707","url":null,"abstract":"An approach to a macroscopic description of a cluster-formation algorithm by neural activities in a random neural network is considered. The activity interaction between clusters of neurons and the network entropy through the medium of the activity parameter x(p) for the input pattern p, are introduced as a system energy. By using the neural state transition rule similar to that in the Boltzmann network and some simple stochastic assumptions, cluster-formation of neurons was simulated. The relations between cluster sizes, or the simulated activity, and the setting activity parameter are shown. The validity of this macroscopic description is also discussed.<<ETX>>","PeriodicalId":211135,"journal":{"name":"[Proceedings] 1991 IEEE International Joint Conference on Neural Networks","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"1991-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117249351","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：