No description available.

No description available.

Thesis (Ph. D.)--University of Wisconsin--Madison, 1963. / Typescript. Vita. eContent provider-neutral record in process. Description based on print version record. Includes bibliographical references.

This paper investigates perceptron-like devices with an eye toward the recognition of simple predicates, such as parity and connectivity. First "multilayer" perceptrons are investigated to little avail. Then much more powerful perceptrons which have feedback are considered, yielding better results. / Science, Faculty of / Mathematics, Department of / Graduate

The problem of character recognition is used as a vehicle for an investigation of the properties of a particular descriptor of planar patterns. The descriptor arises in the work of Connor [4] on the activity of photosensitive receptors connected in a lateral inhibitory network. It results from a mapping of the planar pattern into a periodic function of one variable. The function is found to be a general descriptor of pattern contour which denotes extrema in contour curvature and is independent of pattern orientation. Based on conjectures regarding human perception of the elements of a pattern set representing alphabetic characters, a recognition machine based on a sequential decision strategy is developed and implemented. Results of the application of the strategy to a mixed-font environment show the pattern descriptor to be capable of high recognition rates. However, these anticipated results may be obtained only through the modification of the current operating system. A number of such changes are proposed and their probable effects discussed. / Science, Faculty of / Computer Science, Department of / Graduate

Improvement of specific character recognition systems through the use of contextual constraints has been demonstrated by several researchers in the field. In this thesis, the effectiveness of using contextual constraints to improve the overall performance of a generalized character recognition system is studied experimentally. A computer simulation system designed to facilitate this investigation for both theoretical and physical character recognition system models is described. A suitable character recognition system model is developed. A number of sequential decision algorithms based on Bayes' optimum decision rule and using different order contextual constraints are implemented as the decision element of the character recognition process. The effect on the performance of the contextual decision algorithms of varying the source data, as well as varying the quality and characterization of the feature extraction system with classifier have been demonstrated. The experimental results for the practical character recognition system model confirmed the findings of others that significant improvement in overall system performance can be achieved through the application of contextual constraints. It was shown that the improvement was directly proportional to the order of contextual information used for low order statistics and that the improvement was most pronounced when these statistics were representative of the source material, the feature extraction system with classifier had a high recognition accuracy initially and its performance was characterized by mistaking the most frequently occurring characters more often than the infrequently occurring characters. It was demonstrated that (1) computer simulation is a useful tool in testing hypotheses concerning the application of contextual constraints to character recognition, (2) qualitative predictions based on theoretical character recognition system models are applicable to practical character recognition systems, and (3) the improvement which can be obtained with the application of contextual constraints to character recognition is sufficient to make machine editing of text competitive with clerical editing. / Applied Science, Faculty of / Electrical and Computer Engineering, Department of / Graduate

No description available.

No description available.

Neural networks are models which have been developed to simulate the anatomy of the nervous system. The connection between the elements of these networks, the so called artificial neurons, is similar to the connection between the biological neurons. In developing neural networks people are trying to create systems which have the same computational and communication properties as the brain. On the basis of the things we know from neurophysiology the first models for the neural networks are developed. One of these networks was the perceptron, which is one of the most used neural networks. In this thesis we'll study this special neural network. When the input vectors of the perceptron can be linearly separated into two categories, this network can be trained to correctly classify these input vectors. However in most practical cases the linearly separability assumption isn't satisfied. That's why the main part of this study is devoted to the case where the input vectors aren't linearly separable. / Graduation date: 1995

No description available.