Dynamic Analysis of Feedforward Neural Networks Using Simulated and Measured Data
Abstract
An environment is developed for the study of dynamic changes in patterns of weight and node values for artificial neural networks. Graphic representations of neural network internal states are displayed using a high resolution video terminal. Patterns of node firings and changes in weight vectors are displayed to provide insight during training. Four pattern recognition problems are applied to four types of artificial neural networks. Using simulated data, a simple disjoint region classification problem is developed and examined using a Kohonen net and a multilayer feedforward back propagation (MFB) network. A MFB neural network is also used to simulate a Fourier filter. Using a Kohonen net, a MFB, a counterpropagation and a hybrid network, data measured from infrared and laser radar imagery of military vehicles is analyzed. The accuracy and training times for a MFB net and a Hybrid net are compared using an ambiguous decision region problem. Each classification problem is examined and compared to classical, nearest neighbor pattern recognition techniques. Using dynamic analysis, neural network is developed using Kohonen training rules for the first hidden layer followed by one or two hidden layers using standard back propagation rules for training. Advantage of the hybrid network is shown for classification problems involving anomalies characteristic of measured data. The Hybrid network requires less training and fewer interconnections than MFB when classifications involves ambiguous decision regions. Theses.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1988
- Accession Number
- ADA202573
Entities
People
- Gregory L. Tarr
Organizations
- Air Force Institute of Technology