Extracting algorithms from pattern classification neural networks

doi:10.1016/S0893-6080(05)80117-0

Neural Networks

Volume 6, Issue 5, 1993, Pages 729-735

https://doi.org/10.1016/S0893-6080(05)80117-0 Get rights and content

Abstract

It has been shown that a pattern classification neural network can be synthesized in three layers (one hidden layer) if each class is separated into a single convex region by hyperplanes and in four layers if not. Based upon this, extraction of a pattern classification algorithm from a network is discussed, namely, first separation hyperplanes are extracted from a network trained by the backpropagation algorithm, and then the weights corresponding to the separation hyperplanes and those connected to the output neurons are successively tuned to improve the generalization ability of the network. Classification is made according to which sides of the hyperplanes a test datum (or the second layer output of a test datum for four layers) is on. Finally, weights of the neural networks, generated by the backpropagation algorithm for a number recognition system, are tuned and classification algorithms extracted.

References (11)

FunahashiK.
On the approximate realization of continuous mapping by neural networks
Neural Networks
(1989)
HornikK. et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989)
AbeS. et al.
How neural networks for pattern recognition can be synthesized
ChesterD.L.
Why two hidden layers are better than one
KayamaM. et al.
Constructing optimal neural networks by linear regression analysis

There are more references available in the full text version of this article.

Cited by (11)

Classification of molecular structure images by using ANN, RF, LBP, HOG, and size reduction methods for early stomach cancer detection
2018, Journal of Molecular Structure
Citation Excerpt :
When analyzing with an artificial neural network, it is not necessary to determine a model form. Instead, the model form is based on the nature of the data given to the input of the artificial neural network [5,1]. The purpose of this article is to assist doctors with a computer-assisted program to determine if molecular structures are cancerous or normal.
Patients who die from stomach cancer are still present. Early diagnosis is crucial in reducing the mortality rate of cancer patients. Therefore, computer aided methods have been developed for early detection in this article. Stomach cancer images were obtained from Fırat University Medical Faculty Pathology Department. The Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG) features of these images are calculated. At the same time, Sammon mapping, Stochastic Neighbor Embedding (SNE), Isomap, Classical multidimensional scaling (MDS), Local Linear Embedding (LLE), Linear Discriminant Analysis (LDA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Laplacian Eigenmaps methods are used for dimensional the reduction of the features. The high dimension of these features has been reduced to lower dimensions using dimensional reduction methods. Artificial neural networks (ANN) and Random Forest (RF) classifiers were used to classify stomach cancer images with these new lower feature sizes. New medical systems have developed to measure the effects of these dimensions by obtaining features in different dimensional with dimensional reduction methods. When all the methods developed are compared, it has been found that the best accuracy results are obtained with LBP_MDS_ANN and LBP_LLE_ANN methods.
Lesion size quantification in SPECT using an artificial neural network classification approach
1995, Computers and Biomedical Research
An artificial neural network (ANN) has been developed to determine the size of lesions detected in single photon emission computed tomographic images. The network is the Learning Vector Quantizer and is trained to perform size quantification based on image neighborhoods extracted around the lesions. The ANN is compared to the optimal, Bayesian algorithm developed to perform the same task using the unreconstructed, projection data. The performance of the neural network is evaluated at two different noise levels. The Bayesian algorithm provides the upper bound for size quantification performance against which the ANN is compared. In the ideal case where the Bayesian algorithm has explicit knowledge of the underlying distributions, its performance is superior to that of the neural network. However, in the more realistic case where the distributions need to be estimated from the same learning sample the ANN was trained on, the two algorithms have comparable performances.
Estimating time series soil moisture by applying recurrent nonlinear autoregressive neural networks to passive microwave data over the Heihe River Basin, China
2017, Remote Sensing
The fault diagnosis of garbage crusher based on ant colony algorithm and neural network
2009, 3rd International Conference on Genetic and Evolutionary Computing, WGEC 2009
Study on intelligence fault diagnosis system for electronic controlled engine
2009, 3rd International Symposium on Intelligent Information Technology Application, IITA 2009
Neuro-fuzzy pattern classification model with rule extraction based on supervised learning
2002, ICONIP 2002 - Proceedings of the 9th International Conference on Neural Information Processing: Computational Intelligence for the E-Age

View all citing articles on Scopus

View full text

Original ContributionExtracting algorithms from pattern classification neural networks

Abstract

Neural Networks

Neural Networks

How neural networks for pattern recognition can be synthesized

Why two hidden layers are better than one

Constructing optimal neural networks by linear regression analysis

Original Contribution
Extracting algorithms from pattern classification neural networks