A backpropagation network for classifying auditory brainstem evoked potentials: input level biasing, temporal and spectral inputs and learning patterns

Abstract

Summary form only given, as follows. The results of an investigation conducted to examine the effects of various input data forms on learning of a neural network for classifying auditory evoked potentials are presented. The long-term objective is to use the classification in an automated device for hearing threshold testing. Feedforward multilayered neural networks trained with the backpropagation method are used. The effects of presenting the data to the neural network in various temporal and spectral modes are explored. Results indicate that temporal and spectral information complement one another and increase performance when used together. Learning curves and dot graphs as they are used in this study may reveal network learning strategies. The nature of such learning patterns found in this study is discussed.<>