Pub Date : 1992-06-07DOI: 10.1109/IJCNN.1992.287090
G. Carpenter, S. Grossberg, K. Iizuka
The authors compare the performance of fuzzy ARTMAP with that of learned vector quantization and backpropagation on a handwritten character recognition task. Training with fuzzy ARTMAP to a fixed criterion used many fewer epochs. Voting with fuzzy ARTMAP yielded the highest recognition rates.<>
{"title":"Comparative performance measures of fuzzy ARTMAP, learned vector quantization, and back propagation for handwritten character recognition","authors":"G. Carpenter, S. Grossberg, K. Iizuka","doi":"10.1109/IJCNN.1992.287090","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.287090","url":null,"abstract":"The authors compare the performance of fuzzy ARTMAP with that of learned vector quantization and backpropagation on a handwritten character recognition task. Training with fuzzy ARTMAP to a fixed criterion used many fewer epochs. Voting with fuzzy ARTMAP yielded the highest recognition rates.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115394441","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227082
H. Paugam-Moisy
A block-gradient algorithm is defined, where the weight matrix is updated after every presentation of a block of b examples each. Total and stochastic gradients are included in the block-gradient algorithm, for particular values of b. Experimental laws are stated on the speed of convergence, according to the block size. The first law indicates that an adaptive learning rate has to respect an exponential decreasing function of the number of examples presented between two successive weight updates. The second law states that, with an adaptive learning rate value, the number of epochs grows linearly with the size of the exemplar blocks. The last one shows how the number of epochs for reaching a given level of performance depends on the learning rate.<>
{"title":"On the convergence of a block-gradient algorithm for back-propagation learning","authors":"H. Paugam-Moisy","doi":"10.1109/IJCNN.1992.227082","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227082","url":null,"abstract":"A block-gradient algorithm is defined, where the weight matrix is updated after every presentation of a block of b examples each. Total and stochastic gradients are included in the block-gradient algorithm, for particular values of b. Experimental laws are stated on the speed of convergence, according to the block size. The first law indicates that an adaptive learning rate has to respect an exponential decreasing function of the number of examples presented between two successive weight updates. The second law states that, with an adaptive learning rate value, the number of epochs grows linearly with the size of the exemplar blocks. The last one shows how the number of epochs for reaching a given level of performance depends on the learning rate.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115705023","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227071
C.-L. Chen, R. S. Nutter
Based on the idea of using heterogeneous processing units (PUs) in a network, a variation of the backpropagation (BP) learning algorithm is presented. Three parameters, which are adjustable like connection weights, are incorporated into each PU to increase its autonomous capability by enhancing the output function. The extended BP learning algorithm thus is developed by updating the three parameters as well as connection weights. The extended BP is intended not only to improve the learning speed, but also to reduce the occurrence of local minima. The algorithm has been intensively tested on the XOR problem. By carefully choosing learning rates, results show that the extended BP appears to have advantages over the standard BP in terms of faster learning speed and fewer local minima.<>
{"title":"An extended back-propagation learning algorithm by using heterogeneous processing units","authors":"C.-L. Chen, R. S. Nutter","doi":"10.1109/IJCNN.1992.227071","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227071","url":null,"abstract":"Based on the idea of using heterogeneous processing units (PUs) in a network, a variation of the backpropagation (BP) learning algorithm is presented. Three parameters, which are adjustable like connection weights, are incorporated into each PU to increase its autonomous capability by enhancing the output function. The extended BP learning algorithm thus is developed by updating the three parameters as well as connection weights. The extended BP is intended not only to improve the learning speed, but also to reduce the occurrence of local minima. The algorithm has been intensively tested on the XOR problem. By carefully choosing learning rates, results show that the extended BP appears to have advantages over the standard BP in terms of faster learning speed and fewer local minima.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"16 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121826530","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.287097
Hee-Sook Choi, K. Lee, Yung Hwan Kim, Won Don Lee
A method that minimizes the energy function on the variation not only of weight but also of temperature for the Coulomb energy network (CEN) is proposed. The proposed method is compared with the traditional learning method using only weight variation. It is shown that learning is done more efficiently and accurately with the proposed method. Since weight and temperature can be learned in parallel, the speed of learning might be doubled if appropriate hardware support is provided. The concept of the distance is used to solve the linearly nonseparable classification problem, which cannot be solved in the traditional supervised CEN.<>
{"title":"Learning of the Coulomb energy network on the variation of the temperature","authors":"Hee-Sook Choi, K. Lee, Yung Hwan Kim, Won Don Lee","doi":"10.1109/IJCNN.1992.287097","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.287097","url":null,"abstract":"A method that minimizes the energy function on the variation not only of weight but also of temperature for the Coulomb energy network (CEN) is proposed. The proposed method is compared with the traditional learning method using only weight variation. It is shown that learning is done more efficiently and accurately with the proposed method. Since weight and temperature can be learned in parallel, the speed of learning might be doubled if appropriate hardware support is provided. The concept of the distance is used to solve the linearly nonseparable classification problem, which cannot be solved in the traditional supervised CEN.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117052793","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227270
K. R. Miller, P. Zunde
The Hopfield-Tank optimization network has been applied to the model-image matching problem in computer vision using a graph matching formulation. However, the network has been criticized for unreliable convergence to feasible solutions and for poor solution quality, and the graph matching formulation is unable to represent matching problems with multiple object types, and multiple relations, and high-order relations. The Hopfield-Tank network dynamics is generalized to provide a basis for reliable convergence to feasible solutions, for finding high-quality solutions, and for solving a broad class of optimization problems. The extensions include a new technique called attention-shifting, the introduction of high-order connections in the network, and relaxation of the unit hypercube restriction.<>
{"title":"High-order attention-shifting networks for relational structure matching","authors":"K. R. Miller, P. Zunde","doi":"10.1109/IJCNN.1992.227270","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227270","url":null,"abstract":"The Hopfield-Tank optimization network has been applied to the model-image matching problem in computer vision using a graph matching formulation. However, the network has been criticized for unreliable convergence to feasible solutions and for poor solution quality, and the graph matching formulation is unable to represent matching problems with multiple object types, and multiple relations, and high-order relations. The Hopfield-Tank network dynamics is generalized to provide a basis for reliable convergence to feasible solutions, for finding high-quality solutions, and for solving a broad class of optimization problems. The extensions include a new technique called attention-shifting, the introduction of high-order connections in the network, and relaxation of the unit hypercube restriction.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"60 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120961936","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.287143
J. G. Elias, H.-H. Chu, S. M. Meshreki
The silicon implementation of an artificial passive dendritic tree which can be used to process and classify dynamic signals is described. The electrical circuit architecture is modeled after complex neurons in the vertebrate brain which have spatially extensive dendritic tree structures that support large numbers of synapses. The circuit is primarily analog and, as in the biological model system, is virtually immune to process variations and other factors which often plague more conventional circuits. The nonlinear circuit is sensitive to both temporal and spatial signal characteristics but does not make use of the conventional neural network concept of weights, and as such does not use multipliers, adders, or other complex computational devices. As in biological neuronal circuits, a high degree of local connectivity is required. However, unlike biology, multiplexing of connections is done to reduce the number of conductors to a reasonable level for standard packages.<>
{"title":"Silicon implementation of an artificial dendritic tree","authors":"J. G. Elias, H.-H. Chu, S. M. Meshreki","doi":"10.1109/IJCNN.1992.287143","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.287143","url":null,"abstract":"The silicon implementation of an artificial passive dendritic tree which can be used to process and classify dynamic signals is described. The electrical circuit architecture is modeled after complex neurons in the vertebrate brain which have spatially extensive dendritic tree structures that support large numbers of synapses. The circuit is primarily analog and, as in the biological model system, is virtually immune to process variations and other factors which often plague more conventional circuits. The nonlinear circuit is sensitive to both temporal and spatial signal characteristics but does not make use of the conventional neural network concept of weights, and as such does not use multipliers, adders, or other complex computational devices. As in biological neuronal circuits, a high degree of local connectivity is required. However, unlike biology, multiplexing of connections is done to reduce the number of conductors to a reasonable level for standard packages.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"42 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124861469","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227058
J. Antrobus, S. Alankar, D. Deacon, W. Ritter
The human auditory system has a neurophysiological component, mismatch negativity (MMN), that automatically registers change over time of a variety of simple auditory features, e.g., loudness, pitch, duration, and spatial location. A neural network automatic auditory orienting (AAO-MMN) model which simulates the MMN response is described. The main assumption of the proposed AAO-MMN model is that the broad range characteristic of MMN is achieved by local inhibition of the nonlocal thalamic sources of distributed neural activation. The model represents this activation source by a single thalamic (T) unit that is always fully active. The second assumption is that the buildup of MMN over several repetitions of the standard stimulus is accomplished by a local cumulative activation function. All the local accumulator neurons inhibit the nonlocal, steady-state, thalamic activation represented by T.<>
{"title":"Auditory orienting: automatic detection of auditory change over brief intervals of time: a neural net model of evoked brain potentials","authors":"J. Antrobus, S. Alankar, D. Deacon, W. Ritter","doi":"10.1109/IJCNN.1992.227058","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227058","url":null,"abstract":"The human auditory system has a neurophysiological component, mismatch negativity (MMN), that automatically registers change over time of a variety of simple auditory features, e.g., loudness, pitch, duration, and spatial location. A neural network automatic auditory orienting (AAO-MMN) model which simulates the MMN response is described. The main assumption of the proposed AAO-MMN model is that the broad range characteristic of MMN is achieved by local inhibition of the nonlocal thalamic sources of distributed neural activation. The model represents this activation source by a single thalamic (T) unit that is always fully active. The second assumption is that the buildup of MMN over several repetitions of the standard stimulus is accomplished by a local cumulative activation function. All the local accumulator neurons inhibit the nonlocal, steady-state, thalamic activation represented by T.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"28 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125088101","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227161
R. Laganière, F. Labrosse, P. Cohen
The authors propose a parallel architecture for computing the 3-D structure of a moving scene from a long image sequence, using a principle known as the incremental rigidity scheme. At each instant an internal model of the 3-D structure is updated, based upon the observations accumulated until that time. The updating process favors rigid transformations but tolerates a limited deviation from rigidity. This deviation eventually leads the internal model to converge towards the actual 3-D structure of the scene. The main advantage of this architecture is its ability to accurately estimate the 3-D structure of the scene at a low computational cost. Testing has been successfully performed on synthetic data as well as real image sequences.<>
{"title":"A parallel network for the computation of structure from long-range motion","authors":"R. Laganière, F. Labrosse, P. Cohen","doi":"10.1109/IJCNN.1992.227161","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227161","url":null,"abstract":"The authors propose a parallel architecture for computing the 3-D structure of a moving scene from a long image sequence, using a principle known as the incremental rigidity scheme. At each instant an internal model of the 3-D structure is updated, based upon the observations accumulated until that time. The updating process favors rigid transformations but tolerates a limited deviation from rigidity. This deviation eventually leads the internal model to converge towards the actual 3-D structure of the scene. The main advantage of this architecture is its ability to accurately estimate the 3-D structure of the scene at a low computational cost. Testing has been successfully performed on synthetic data as well as real image sequences.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"51 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126080983","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227131
J. Steck
Three theorems are presented which establish an upper bound on the magnitude of the weights which guarantees convergence of the network to a stable unique fixed point. It is shown that the bound on the weights is inversely proportional to the product of the number of neurons in the network and the maximum slope of the neuron activation functions. The location of its fixed point is determined by the network architecture, weights, and the external input values. The proofs are constructive, consisting of representing the network as a contraction mapping and then applying the contraction mapping theorem from point set topology. The resulting sufficient conditions for network stability are shown to be general enough to allow the network to have nontrivial fixed points.<>
{"title":"Convergence of recurrent networks as contraction mappings","authors":"J. Steck","doi":"10.1109/IJCNN.1992.227131","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227131","url":null,"abstract":"Three theorems are presented which establish an upper bound on the magnitude of the weights which guarantees convergence of the network to a stable unique fixed point. It is shown that the bound on the weights is inversely proportional to the product of the number of neurons in the network and the maximum slope of the neuron activation functions. The location of its fixed point is determined by the network architecture, weights, and the external input values. The proofs are constructive, consisting of representing the network as a contraction mapping and then applying the contraction mapping theorem from point set topology. The resulting sufficient conditions for network stability are shown to be general enough to allow the network to have nontrivial fixed points.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125350462","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 : 1992-06-07DOI: 10.1109/IJCNN.1992.227157
G. Carpenter, S. Grossberg, G. Lesher
A feedforward neural networks for invariant image preprocessing is proposed that represents the position, orientation, and size of an image figure (where it is) in a multiplexed spatial map. This map is used to generate an invariant representation of the figure that is insensitive to position, orientation, and size for purposes of pattern recognition (what it is). Image recognition is based upon the output from the what channel. A multiscale array of oriented filters, followed by competition between orientations and scales, is used to define the where filter.<>
{"title":"A what-and-where neural network for invariant image preprocessing","authors":"G. Carpenter, S. Grossberg, G. Lesher","doi":"10.1109/IJCNN.1992.227157","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227157","url":null,"abstract":"A feedforward neural networks for invariant image preprocessing is proposed that represents the position, orientation, and size of an image figure (where it is) in a multiplexed spatial map. This map is used to generate an invariant representation of the figure that is insensitive to position, orientation, and size for purposes of pattern recognition (what it is). Image recognition is based upon the output from the what channel. A multiscale array of oriented filters, followed by competition between orientations and scales, is used to define the where filter.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"54 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1992-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126609799","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：