Fortran Based Learning System Using Multilayer Back-Propagation Neural NetworkTechniques

摘要

An interactive computer system which allows the researcher to build an 'optimal'neural network structure quickly, is developed and validated. This system assumes a single hidden layer perceptron structure and uses the back-propagation training technique. The software enables the researcher to quickly define a neural network structure, train the neural network, interrupt training at any point to analyze the status of the current network, re-start training at the interrupted point if desired, and analyze the final network using two-dimensional graphs, three-dimensional graphs, confusion matrices and saliency metrics. A technique for training, testing, and validating various network structures and parameters, using the interactive computer system, is demonstrated. Outputs automatically produced by the system are analyzed in an iterative fashion, resulting in an 'optimal' neural network structure tailored for the specific problem. To validate the system, the technique is applied to two classic classification problems. The first is the two-class XOR problem. The second is the four-class MESH problem. Noise variables are introduced to determine if weight monitoring graphs, saliency metrics and saliency grids can detect them. Three dimensional class activation grids and saliency grids are analyzed to determine class borders of the two problems. Results of the validation process showed that this interactive computer system is a valuable tool in determining an optimal network structure, given a specific problem. Neural networks, Pattern recognition, Back-propagation, Learning system, Perceptron.