부산대학교 도서관

Today's neurocomputation usually is based on complete software emulation and is therefore often called neurosimulation. Inputs, outputs, neurons, synapses and weights are implemented in software. The neurosimulator FAUN (Fast Approximation with Universal Neural networks) enables supervised learning with 3- and 4-layered perceptrons and also radial basis functions. A FAUN user has to provide patterns, i.e. input-output pairs explaining a mathematical relation. Then artificial neural networks (ANN) are trained to learn the relation with a black-box approach. A well trained ANN reasonably interpolates and extrapolates between the patterns (generalization). This discussion paper shows in detail how FAUN works and gives several examples of use.