An environment is developed for the study of artificial neural networks. Graphic representations of neural network internal states and interconnection weights are displayed using a high resolution video terminal. Several pattern recognition problems are examined using a variety of training paradigms. The environment is designed for graphical analysis on a Silicon Graphics IRIS 3130 and an IBM with EGA. Portability without graphics is demonstrated for several Unix based operating systems. The environment includes demonstrations and applications of Kohonen mappings, multi-layer feedforward networks using backpropagation, hybrid (joint supervised/unsupervised) training paradigms, Hopfield associative memory, error surface analysis and network topology analysis. Several pattern recognition problems are tested with the environment to determine suitability for neural network solutions. Applications include radar signal identification, speech recognition, tactical target classification and others.
Neural network development tools at the air force institute of technology wright-patterson AFB
Neuronale Netze als Entwicklungswerkzeuge am Luftwaffeninstitut für Technologie in Wright-Patterson
1989
12 Seiten, 5 Bilder, 15 Quellen
Conference paper
English