Pub Date : 2002-08-06DOI: 10.1109/ICNN.1994.374460
E. V. Keulen, S. Colak, H. Withagen, Hans Hegt
Nowadays, many real world problems need fast processing neural networks to come up with a solution in real time. Therefore hardware implementation becomes indispensable. The problem is then to choose the right chip that is to be used for a particular application. For this, a proper set of hardware performance criteria is needed to be able to compare the performance of neural network chips. The most important criterion is related to the speed a network processes information with a given accuracy. For this a new criterion is proposed. The 'effective number of connection bits' represents the effective accuracy of a chip. The '(effective) connection primitives per second' criterion now provides a new speed criterion normalized to the amount of information value that is processed in a connection. In addition to this we also propose another new criterion called 'reconfigurability number' as a measure for the reconfigurability and size of a chip. Using these criteria gives a much more neutral view of the performance of a neural network chip than the existing conventional criteria, such as 'connections per second'.<>
{"title":"Neural network hardware performance criteria","authors":"E. V. Keulen, S. Colak, H. Withagen, Hans Hegt","doi":"10.1109/ICNN.1994.374460","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374460","url":null,"abstract":"Nowadays, many real world problems need fast processing neural networks to come up with a solution in real time. Therefore hardware implementation becomes indispensable. The problem is then to choose the right chip that is to be used for a particular application. For this, a proper set of hardware performance criteria is needed to be able to compare the performance of neural network chips. The most important criterion is related to the speed a network processes information with a given accuracy. For this a new criterion is proposed. The 'effective number of connection bits' represents the effective accuracy of a chip. The '(effective) connection primitives per second' criterion now provides a new speed criterion normalized to the amount of information value that is processed in a connection. In addition to this we also propose another new criterion called 'reconfigurability number' as a measure for the reconfigurability and size of a chip. Using these criteria gives a much more neutral view of the performance of a neural network chip than the existing conventional criteria, such as 'connections per second'.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132895885","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 : 2002-08-06DOI: 10.1109/ICNN.1994.374883
Jing-long Wu, Y. Nishikawa
This paper proposes a model of binocular fusion based on psychological experimental results and physiological knowledge. Considering the psychological results and the physiological structure, the authors assume that the binocular information is processed by several binocular channels having different spatial characteristics from low spatial frequency to high spatial frequency. In order to examine the mechanism of binocular fusion, the authors construct a five layer neural network model, and train it by the backpropagation learning algorithm using psychological experimental data. After completion of learning, the generalization capability of the network is examined. Further, the response functions of the hidden units have been examined, which suggested that the hidden units have a spatial selective characteristic.<>
{"title":"A neural network model of the binocular fusion in the human vision","authors":"Jing-long Wu, Y. Nishikawa","doi":"10.1109/ICNN.1994.374883","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374883","url":null,"abstract":"This paper proposes a model of binocular fusion based on psychological experimental results and physiological knowledge. Considering the psychological results and the physiological structure, the authors assume that the binocular information is processed by several binocular channels having different spatial characteristics from low spatial frequency to high spatial frequency. In order to examine the mechanism of binocular fusion, the authors construct a five layer neural network model, and train it by the backpropagation learning algorithm using psychological experimental data. After completion of learning, the generalization capability of the network is examined. Further, the response functions of the hidden units have been examined, which suggested that the hidden units have a spatial selective characteristic.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2002-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129092928","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 : 1996-03-01DOI: 10.1109/ICNN.1994.374134
N. Karayiannis
It is argued in this paper that most of the problems associated with the application of existing learning algorithms in complex training tasks can be overcome by using only the input data to determine the role of the hidden units, which form a data compression or a data expansion layer. The initial set of internal representations can be formed through an unsupervised learning process applied before the supervised training algorithm. The synaptic weights that connect the input of the network with the hidden units can be determined through various linear or nonlinear variations of a generalized Hebbian learning rule, known as the Oja's rule. Several experiments indicated that the use of the proposed initialization of the internal representations improves significantly the convergence of various gradient-descent-based algorithms used to perform nontrivial training tasks.<>
{"title":"Accelerating the training of feedforward neural networks using generalized Hebbian rules for initializing the internal representations","authors":"N. Karayiannis","doi":"10.1109/ICNN.1994.374134","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374134","url":null,"abstract":"It is argued in this paper that most of the problems associated with the application of existing learning algorithms in complex training tasks can be overcome by using only the input data to determine the role of the hidden units, which form a data compression or a data expansion layer. The initial set of internal representations can be formed through an unsupervised learning process applied before the supervised training algorithm. The synaptic weights that connect the input of the network with the hidden units can be determined through various linear or nonlinear variations of a generalized Hebbian learning rule, known as the Oja's rule. Several experiments indicated that the use of the proposed initialization of the internal representations improves significantly the convergence of various gradient-descent-based algorithms used to perform nontrivial training tasks.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1996-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115889155","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 : 1995-02-25DOI: 10.1109/ICNN.1994.374153
R. Kamimura, T. Takagi, S. Nakanishi
In this paper, we attempt to show that the information stored in networks must be as small as possible for the improvement of the generalization performance under the condition that the networks can produce targets with appropriate accuracy. The information is defined by the difference between maximum entropy or uncertainty and observed entropy. Borrowing a definition of fuzzy entropy, the uncertainty function is defined for the internal representation and represented by the equation: -/spl upsi//sub i/ log /spl upsi//sub i/-(1-/spl upsi//sub i/) log (1-/spl upsi//sub i/), where /spl upsi//sub i/ is a hidden unit activity. After having formulated an update rule for the minimization of the information, we applied the method to a problem of language acquisition: the inference of the past tense forms of regular verbs. Experimental results confirmed that by our method, the information was significantly decreased and the generalization performance was greatly improved.<>
{"title":"Improving generalization performance by information minimization","authors":"R. Kamimura, T. Takagi, S. Nakanishi","doi":"10.1109/ICNN.1994.374153","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374153","url":null,"abstract":"In this paper, we attempt to show that the information stored in networks must be as small as possible for the improvement of the generalization performance under the condition that the networks can produce targets with appropriate accuracy. The information is defined by the difference between maximum entropy or uncertainty and observed entropy. Borrowing a definition of fuzzy entropy, the uncertainty function is defined for the internal representation and represented by the equation: -/spl upsi//sub i/ log /spl upsi//sub i/-(1-/spl upsi//sub i/) log (1-/spl upsi//sub i/), where /spl upsi//sub i/ is a hidden unit activity. After having formulated an update rule for the minimization of the information, we applied the method to a problem of language acquisition: the inference of the past tense forms of regular verbs. Experimental results confirmed that by our method, the information was significantly decreased and the generalization performance was greatly improved.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1995-02-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122459108","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 : 1994-12-31DOI: 10.1109/ICNN.1994.374640
S. Matsumura, S. Omatu, H. Higasa
In order to develop an efficient driving system for electric vehicle (EV), a testing system using motors has been built to simulate the driving performance of EVs. In the testing system, the PID controller is used to control rotating speed of motor when the EV drives. In this paper, in order to improve the performance of speed control, a neural network is applied to tuning parameters of PI controller. It is shown,through experiments that a neural network can reduce output error effectively while the PI controller parameters are being tuned online.<>
{"title":"Improvement of speed control performance using PID type neurocontroller in an electric vehicle system","authors":"S. Matsumura, S. Omatu, H. Higasa","doi":"10.1109/ICNN.1994.374640","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374640","url":null,"abstract":"In order to develop an efficient driving system for electric vehicle (EV), a testing system using motors has been built to simulate the driving performance of EVs. In the testing system, the PID controller is used to control rotating speed of motor when the EV drives. In this paper, in order to improve the performance of speed control, a neural network is applied to tuning parameters of PI controller. It is shown,through experiments that a neural network can reduce output error effectively while the PI controller parameters are being tuned online.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133145995","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 : 1994-12-01DOI: 10.1109/ICNN.1994.374147
Sheng-Tun Li, Yiwei Chen, E. Leiss
A variety of neural models, especially higher-order networks, are known to be computationally powerful for complex applications. While they have advantages over traditional multilayer perceptrons, the nonuniformity in their network structures and learning algorithms creates practical problems. Thus there is a need for a framework that unifies these various models. This paper presents a novel neuron model, called generalized multi-dendrite product (GMDP) unit. Multilayer feedforward neural networks with GMDP units are shown to be capable of realizing higher-order neural networks. The standard backpropagation learning rule is extended to this neural network. Simulation results show that single layer GMDP networks provide an efficient model for solving general problems on function approximation and pattern classification.<>
{"title":"GMDP: a novel unified neuron model for multilayer feedforward neural networks","authors":"Sheng-Tun Li, Yiwei Chen, E. Leiss","doi":"10.1109/ICNN.1994.374147","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374147","url":null,"abstract":"A variety of neural models, especially higher-order networks, are known to be computationally powerful for complex applications. While they have advantages over traditional multilayer perceptrons, the nonuniformity in their network structures and learning algorithms creates practical problems. Thus there is a need for a framework that unifies these various models. This paper presents a novel neuron model, called generalized multi-dendrite product (GMDP) unit. Multilayer feedforward neural networks with GMDP units are shown to be capable of realizing higher-order neural networks. The standard backpropagation learning rule is extended to this neural network. Simulation results show that single layer GMDP networks provide an efficient model for solving general problems on function approximation and pattern classification.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128186803","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 : 1994-12-01DOI: 10.1109/ICNN.1994.374346
K. Ikeda, Noboru Murata, S. Amari
The more the number of training examples included, the better a learning machine will behave. It is an important to know how fast and how well the behavior is improved. The average prediction error is one of the most popular criteria to evaluate the behavior. We have regarded the machine learning from the point of view of parameter estimation and derived the average prediction error of stochastic dichotomy machines by the information geometrical method.<>
{"title":"Prediction error of stochastic learning machine","authors":"K. Ikeda, Noboru Murata, S. Amari","doi":"10.1109/ICNN.1994.374346","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374346","url":null,"abstract":"The more the number of training examples included, the better a learning machine will behave. It is an important to know how fast and how well the behavior is improved. The average prediction error is one of the most popular criteria to evaluate the behavior. We have regarded the machine learning from the point of view of parameter estimation and derived the average prediction error of stochastic dichotomy machines by the information geometrical method.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"166 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132958022","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 : 1994-12-01DOI: 10.1109/ICNN.1994.374179
V. Kreinovich, O. Sirisaengtaksin, S. Cabrera
Neural networks are universal approximators. For example, it has been proved (K. Hornik et al., 1989) that for every /spl epsiv/>0 an arbitrary continuous function on a compact set can be /spl epsiv/-approximated by a 3-layer neural network. This and other results prove that in principle, any function (e.g., any control) can be implemented by an appropriate neural network. But why neural networks? In addition to neural networks, an arbitrary continuous function can be also approximated by polynomials, etc. What is so special about neural networks that make them preferable approximators? To compare different approximators, one can compare the number of bits that we must store in order to be able to reconstruct a function with a given precision /spl epsiv/. For neural networks, we must store weights and thresholds. For polynomials, we must store coefficients, etc. We consider functions of one variable, and show that for some special neurons (corresponding to wavelets), neural networks are optimal approximators in the sense that they require (asymptotically) the smallest possible number of bits.<>
神经网络是通用逼近器。例如，已经证明(K. Hornik et al.， 1989)对于每一个/spl epsiv/>0，紧集上的任意连续函数都可以用三层神经网络逼近/spl epsiv/-。这一结果和其他结果证明，原则上，任何函数(例如，任何控制)都可以由适当的神经网络实现。但为什么是神经网络呢?除了神经网络，任意连续函数也可以用多项式等逼近。神经网络有什么特别之处使其成为更好的近似器?为了比较不同的近似值，可以比较我们必须存储的比特数，以便能够以给定的精度/spl epsiv/重建函数。对于神经网络，我们必须存储权值和阈值。对于多项式，我们必须存储系数，等等。我们考虑单变量的函数，并表明对于一些特殊的神经元(对应于小波)，神经网络是最优逼近器，因为它们(渐近地)需要尽可能少的比特数。
{"title":"Wavelet neural networks are asymptotically optimal approximators for functions of one variable","authors":"V. Kreinovich, O. Sirisaengtaksin, S. Cabrera","doi":"10.1109/ICNN.1994.374179","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374179","url":null,"abstract":"Neural networks are universal approximators. For example, it has been proved (K. Hornik et al., 1989) that for every /spl epsiv/>0 an arbitrary continuous function on a compact set can be /spl epsiv/-approximated by a 3-layer neural network. This and other results prove that in principle, any function (e.g., any control) can be implemented by an appropriate neural network. But why neural networks? In addition to neural networks, an arbitrary continuous function can be also approximated by polynomials, etc. What is so special about neural networks that make them preferable approximators? To compare different approximators, one can compare the number of bits that we must store in order to be able to reconstruct a function with a given precision /spl epsiv/. For neural networks, we must store weights and thresholds. For polynomials, we must store coefficients, etc. We consider functions of one variable, and show that for some special neurons (corresponding to wavelets), neural networks are optimal approximators in the sense that they require (asymptotically) the smallest possible number of bits.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115443218","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 : 1994-12-01DOI: 10.1109/ICNN.1994.374887
B. Zheng, W. Qian, L. Clarke
A novel multistage neural network (MSNN) is proposed for locating and classification of micro-calcification in digital mammography. Backpropagation (BP) with Kalman filtering (KF) is used for training the MSNN. A new nonlinear decision method is proposed to improve the performance of the classification. The experimental results show that the sensitivity of this classification/detection is 100% with the false positive detection rate of less than 1 micro-calcification clusters (MCCs) per image. The proposed methods are automatic or operator independent and provide realistic image processing times as required for breast cancer screening programs. Full clinical analysis is planned using large databases.<>
{"title":"Multistage neural network for pattern recognition in mammogram screening","authors":"B. Zheng, W. Qian, L. Clarke","doi":"10.1109/ICNN.1994.374887","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374887","url":null,"abstract":"A novel multistage neural network (MSNN) is proposed for locating and classification of micro-calcification in digital mammography. Backpropagation (BP) with Kalman filtering (KF) is used for training the MSNN. A new nonlinear decision method is proposed to improve the performance of the classification. The experimental results show that the sensitivity of this classification/detection is 100% with the false positive detection rate of less than 1 micro-calcification clusters (MCCs) per image. The proposed methods are automatic or operator independent and provide realistic image processing times as required for breast cancer screening programs. Full clinical analysis is planned using large databases.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123891346","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 : 1994-12-01DOI: 10.1109/ICNN.1994.374891
Y. Hu, K. Ashenayi, R. Veltri, G. O'Dowd, G. Miller, R. Hurst, R. Bonner
We report the performances of cancer cell classification by using supervised and unsupervised learning techniques. A single hidden layer feedforward NN with error back-propagation training is adopted for supervised learning, and c-means clustering methods, fuzzy and nonfuzzy, are used for unsupervised learning. Network configurations with various activation functions, namely sigmoid, sinusoid and gaussian, are studied. A set of features, including cell size, average intensity, texture, shape factor and pgDNA are selected as the input for the network. These features, in particular the texture information, are shown to be very effective in capturing the discriminate information in cancer cells. It is found, based on the data from 467 cell images from six cases, the neural network approach achieves a classification rate of 96.9% while fuzzy c-means scores 76.5%.<>
{"title":"A comparison of neural network and fuzzy c-means methods in bladder cancer cell classification","authors":"Y. Hu, K. Ashenayi, R. Veltri, G. O'Dowd, G. Miller, R. Hurst, R. Bonner","doi":"10.1109/ICNN.1994.374891","DOIUrl":"https://doi.org/10.1109/ICNN.1994.374891","url":null,"abstract":"We report the performances of cancer cell classification by using supervised and unsupervised learning techniques. A single hidden layer feedforward NN with error back-propagation training is adopted for supervised learning, and c-means clustering methods, fuzzy and nonfuzzy, are used for unsupervised learning. Network configurations with various activation functions, namely sigmoid, sinusoid and gaussian, are studied. A set of features, including cell size, average intensity, texture, shape factor and pgDNA are selected as the input for the network. These features, in particular the texture information, are shown to be very effective in capturing the discriminate information in cancer cells. It is found, based on the data from 467 cell images from six cases, the neural network approach achieves a classification rate of 96.9% while fuzzy c-means scores 76.5%.<<ETX>>","PeriodicalId":209128,"journal":{"name":"Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94)","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1994-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121163064","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：