Pub Date : 1992-06-07DOI: 10.1109/IJCNN.1992.287164
J. Hou, F. Salam
The authors present a model for recurrent artificial neural networks which can store any number of any prespecified patterns as energy local minima. Therefore, all the prespecified patterns can be stored and retrieved. The authors summarize the model's stability properties. They then give two examples, showing how this model can be used in image recognition and association.<>
{"title":"A product-of-norms model for recurrent neural networks","authors":"J. Hou, F. Salam","doi":"10.1109/IJCNN.1992.287164","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.287164","url":null,"abstract":"The authors present a model for recurrent artificial neural networks which can store any number of any prespecified patterns as energy local minima. Therefore, all the prespecified patterns can be stored and retrieved. The authors summarize the model's stability properties. They then give two examples, showing how this model can be used in image recognition and association.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"35 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":"127872418","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.1992.227283
Y. Peng
The author presents an unsupervised method to learn probabilities of random events. Learning is done by letting variables adaptively respond to positive and negative environmental stimuli. The basic learning rule is applied to learn prior and conditional probabilities for causal networks. By combining with a stochastic factor, this method is extended to learn probabilities of hidden causations, a type of event important in modeling causal relationships. In contrast to many existing neural network learning paradigms, probabilistic knowledge learned by this method is independent of any particular type of task. This method is especially suited for acquiring and updating knowledge in systems based on traditional artificial intelligence representation techniques.<>
{"title":"Learning probabilities for causal networks","authors":"Y. Peng","doi":"10.1109/IJCNN.1992.227283","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227283","url":null,"abstract":"The author presents an unsupervised method to learn probabilities of random events. Learning is done by letting variables adaptively respond to positive and negative environmental stimuli. The basic learning rule is applied to learn prior and conditional probabilities for causal networks. By combining with a stochastic factor, this method is extended to learn probabilities of hidden causations, a type of event important in modeling causal relationships. In contrast to many existing neural network learning paradigms, probabilistic knowledge learned by this method is independent of any particular type of task. This method is especially suited for acquiring and updating knowledge in systems based on traditional artificial intelligence representation techniques.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"521 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":"115352150","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.1992.227076
C. Ji
A general relationship is developed between the two sharp transition points, the statistical capacity which represents the memorization, and the universal sample bound for generalization, for a network composed of random samples drawn from a specific class of distributions. This relationship indicates that generalization happens after memorization. It is shown through one example that the sample complexity needed for generalization can coincide with the capacity point. For the worst case, the sample complexity for generalization can be on the order of the distribution-free bound, whereas, for a more structured case, it can be smaller than the worst case bound. The analysis sheds light on why in practice the number of samples needed for generalization can be smaller than the bound given in term of the VC-dimension.<>
{"title":"Is the distribution-free sample bound for generalization tight?","authors":"C. Ji","doi":"10.1109/IJCNN.1992.227076","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227076","url":null,"abstract":"A general relationship is developed between the two sharp transition points, the statistical capacity which represents the memorization, and the universal sample bound for generalization, for a network composed of random samples drawn from a specific class of distributions. This relationship indicates that generalization happens after memorization. It is shown through one example that the sample complexity needed for generalization can coincide with the capacity point. For the worst case, the sample complexity for generalization can be on the order of the distribution-free bound, whereas, for a more structured case, it can be smaller than the worst case bound. The analysis sheds light on why in practice the number of samples needed for generalization can be smaller than the bound given in term of the VC-dimension.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"12 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":"115400314","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.1992.227148
R. A. Clinkenbeard, X. Feng
A hybrid unsupervised neural network and fuzzy logic approach is presented to achieve the primary goal of software categorization and feature interpretation. This method permits new software applications to be evaluated quickly for capacity planning and project management purposes. Fuzzy logic techniques were successfully applied to interpret the internal structure of the trained network, leading to an understanding of which application attributes most clearly distinguish the resulting categories. The resulting fuzzy membership functions can be used as inputs to subsequent analysis. These techniques can derive useful categories based on broad, external attributes of the software. This makes the technique useful to users of off-the-shelf software or to developers in the early stages of program specification. Experiments explicitly demonstrated the advantages of this method.<>
{"title":"An unsupervised learning and fuzzy logic approach for software category identification and capacity planning","authors":"R. A. Clinkenbeard, X. Feng","doi":"10.1109/IJCNN.1992.227148","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227148","url":null,"abstract":"A hybrid unsupervised neural network and fuzzy logic approach is presented to achieve the primary goal of software categorization and feature interpretation. This method permits new software applications to be evaluated quickly for capacity planning and project management purposes. Fuzzy logic techniques were successfully applied to interpret the internal structure of the trained network, leading to an understanding of which application attributes most clearly distinguish the resulting categories. The resulting fuzzy membership functions can be used as inputs to subsequent analysis. These techniques can derive useful categories based on broad, external attributes of the software. This makes the technique useful to users of off-the-shelf software or to developers in the early stages of program specification. Experiments explicitly demonstrated the advantages of this method.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"84 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":"115443080","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.1992.227075
D. McMichael
BARTIN (Bayesian real-time network) is a general structure for learning Bayesian minimum-risk decision schemes. It comprises two unspecified supervised learning nets and associated elements. The structure allows separate prior compensation and risk minimization and is thus able to learn Bayesian minimum-risk decision schemes accurately from training data and priors alone. The design provides a new mechanism (the prior compensator) for correcting for discrepancies between class probabilities in training and recall. The same mechanism can be adapted to bias output decisions. The general structure of BARTIN is described and the enumerative and Gaussian specific form are presented. The enumerative form of BARTIN was applied to a visual inspection problem in comparison with the multilayer perceptron.<>
{"title":"A general scheme for minimising Bayes risk and incorporating priors applicable to supervised learning systems","authors":"D. McMichael","doi":"10.1109/IJCNN.1992.227075","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227075","url":null,"abstract":"BARTIN (Bayesian real-time network) is a general structure for learning Bayesian minimum-risk decision schemes. It comprises two unspecified supervised learning nets and associated elements. The structure allows separate prior compensation and risk minimization and is thus able to learn Bayesian minimum-risk decision schemes accurately from training data and priors alone. The design provides a new mechanism (the prior compensator) for correcting for discrepancies between class probabilities in training and recall. The same mechanism can be adapted to bias output decisions. The general structure of BARTIN is described and the enumerative and Gaussian specific form are presented. The enumerative form of BARTIN was applied to a visual inspection problem in comparison with the multilayer perceptron.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"31 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":"115773559","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.1992.227069
G. Linhart, G. Dorffner
A proposal by M. B. Reid et al. (1989) to improve the efficiency of higher-order neural networks was built into a pattern recognition system that autonomously learns to categorize and recognize patterns independently of their position in an input image. It does this by combining higher-order with first-order networks and the mechanisms known from ART. Its recognition is based on a 16*16 pixel input which contains a section of the image found by a separate centering mechanism. With this system position invariant recognition can be implemented efficiently, while combining all the advantages of the subsystems.<>
M. B. Reid等人(1989)提出了一个提高高阶神经网络效率的建议，该建议被构建到一个模式识别系统中，该系统可以自主学习分类和识别模式，而不依赖于模式在输入图像中的位置。它通过结合高阶网络和一阶网络以及ART中已知的机制来做到这一点。它的识别是基于一个16*16像素的输入，其中包含由一个单独的定心机制找到的图像的一部分。该系统结合了各子系统的优点，可以有效地实现位置不变识别。
{"title":"A self-learning visual pattern explorer and recognizer using a higher order neural network","authors":"G. Linhart, G. Dorffner","doi":"10.1109/IJCNN.1992.227069","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227069","url":null,"abstract":"A proposal by M. B. Reid et al. (1989) to improve the efficiency of higher-order neural networks was built into a pattern recognition system that autonomously learns to categorize and recognize patterns independently of their position in an input image. It does this by combining higher-order with first-order networks and the mechanisms known from ART. Its recognition is based on a 16*16 pixel input which contains a section of the image found by a separate centering mechanism. With this system position invariant recognition can be implemented efficiently, while combining all the advantages of the subsystems.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"22 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":"124156412","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.1992.287089
N. Karunanithi, L. D. Whitley
The authors present an adaptive modeling approach based on connectionist networks and demonstrate how both feedforward and recurrent networks and various training regimes can be applied to predict software reliability. They make an empirical comparison between this new approach and five well-known software reliability growth prediction models using data sets from 14 different software projects. The results presented suggest that connectionist networks adapt well to different data sets and exhibit better overall long-term predictive accuracy than the analytic models. This observation is true not only for the aggregate data, but for each individual item of data as well. The connectionist approach offers a distinct advantage for software reliability modeling in that the model development is automatic if one uses a training algorithm such as the cascade correlation. Two important characteristics of connectionist models are easy construction of appropriate models and good adaptability towards different data sets (i.e., different software projects).<>
{"title":"Prediction of software reliability using feedforward and recurrent neural nets","authors":"N. Karunanithi, L. D. Whitley","doi":"10.1109/IJCNN.1992.287089","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.287089","url":null,"abstract":"The authors present an adaptive modeling approach based on connectionist networks and demonstrate how both feedforward and recurrent networks and various training regimes can be applied to predict software reliability. They make an empirical comparison between this new approach and five well-known software reliability growth prediction models using data sets from 14 different software projects. The results presented suggest that connectionist networks adapt well to different data sets and exhibit better overall long-term predictive accuracy than the analytic models. This observation is true not only for the aggregate data, but for each individual item of data as well. The connectionist approach offers a distinct advantage for software reliability modeling in that the model development is automatic if one uses a training algorithm such as the cascade correlation. Two important characteristics of connectionist models are easy construction of appropriate models and good adaptability towards different data sets (i.e., different software projects).<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"23 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":"124239558","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.1992.227120
M. Nakao, K. Watanabe, T. Takahashi, Y. Mizutani, M. Yamamoto
It was found that single neuronal activities in various regions in the brain commonly exhibit the distinct dynamics transition from the white to the a/f spectral profiles during the sleep cycle in cats. The dynamics transition was simulated by using a symmetrically connected neural network model including a globally applied inhibitory input. The structure of the network attractor was suggested to vary in association with the change in inhibitory level. To examine the robustness of the dynamics transition, the symmetry network structure is extended to the asymmetrically connected network model. This asymmetricity follows the rule which approximately reflects the characteristics of synaptic contacts between neurons. Computer simulations showed that the inhibitory input could change the neuronal dynamics from the white to the 1/f profiles under more realistic situations. The geometry of the network attractor realizing the dynamics transition is discussed.<>
{"title":"Structural properties of network attractor associated with neuronal dynamics transition","authors":"M. Nakao, K. Watanabe, T. Takahashi, Y. Mizutani, M. Yamamoto","doi":"10.1109/IJCNN.1992.227120","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227120","url":null,"abstract":"It was found that single neuronal activities in various regions in the brain commonly exhibit the distinct dynamics transition from the white to the a/f spectral profiles during the sleep cycle in cats. The dynamics transition was simulated by using a symmetrically connected neural network model including a globally applied inhibitory input. The structure of the network attractor was suggested to vary in association with the change in inhibitory level. To examine the robustness of the dynamics transition, the symmetry network structure is extended to the asymmetrically connected network model. This asymmetricity follows the rule which approximately reflects the characteristics of synaptic contacts between neurons. Computer simulations showed that the inhibitory input could change the neuronal dynamics from the white to the 1/f profiles under more realistic situations. The geometry of the network attractor realizing the dynamics transition is discussed.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"411 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":"124388669","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.1992.227159
A. Cerrato, G. Parodi, R. Zunino
An associative architecture for mapping input images into a set of predefined bit patterns (messages) is described. The running general methodology exploits memory content-addressability to perform robust vision tasks. A noiselike coding associative memory works out message samples from input images, while a superimposed feedforward network filters out memory crosstalk and provides clean messages patterns. The integrated structure combines the generalization power of neural networks with the massive processing capability of associative memories. Tests have involved image sets which stress the system's discrimination efficacy. Experimental results confirmed the system's robustness and flexibility. The overall structure can be regarded as a general domain-independent method for visual stimulus-response mapping.<>
{"title":"An integrated associative structure for vision","authors":"A. Cerrato, G. Parodi, R. Zunino","doi":"10.1109/IJCNN.1992.227159","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227159","url":null,"abstract":"An associative architecture for mapping input images into a set of predefined bit patterns (messages) is described. The running general methodology exploits memory content-addressability to perform robust vision tasks. A noiselike coding associative memory works out message samples from input images, while a superimposed feedforward network filters out memory crosstalk and provides clean messages patterns. The integrated structure combines the generalization power of neural networks with the massive processing capability of associative memories. Tests have involved image sets which stress the system's discrimination efficacy. Experimental results confirmed the system's robustness and flexibility. The overall structure can be regarded as a general domain-independent method for visual stimulus-response mapping.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"11 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":"114448210","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.1992.227085
S. J. Huang, S. N. Koh, H. K. Tang
The use of Newton's method with dynamic error control as a training algorithm for the backpropagation (BP) neural network is considered. Theoretically, it can be proved that Newton's method is convergent in the second-order while the most widely used steepest-descent method is convergent in the first-order. This suggests that Newton's method might be a faster algorithm for the BP network. The updating equations of the two methods are analyzed in detail to extract some important properties with reference to the error surface characteristics. The common benchmark XOR problem is used to compare the performance of the methods.<>
{"title":"Training algorithm based on Newton's method with dynamic error control","authors":"S. J. Huang, S. N. Koh, H. K. Tang","doi":"10.1109/IJCNN.1992.227085","DOIUrl":"https://doi.org/10.1109/IJCNN.1992.227085","url":null,"abstract":"The use of Newton's method with dynamic error control as a training algorithm for the backpropagation (BP) neural network is considered. Theoretically, it can be proved that Newton's method is convergent in the second-order while the most widely used steepest-descent method is convergent in the first-order. This suggests that Newton's method might be a faster algorithm for the BP network. The updating equations of the two methods are analyzed in detail to extract some important properties with reference to the error surface characteristics. The common benchmark XOR problem is used to compare the performance of the methods.<<ETX>>","PeriodicalId":286849,"journal":{"name":"[Proceedings 1992] IJCNN International Joint Conference on Neural Networks","volume":"120 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":"117298929","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：