Simulation of learning and transfer in profile analysis by artificial neural networks
Jeffrey L. Harpster, James H. Howard, Maxwell H. Miller · The Journal of the Acoustical Society of America · 1990
Two aspects of profile analysis previously investigated in human listeners by Green and his associates were simulated by artificial neural networks (ANNs). First, the time course of learning was compared in the neural networks and the human listeners. Both had learning curves that showed a rapid initial improvement followed by gradual leveling to asymptote. Several values of the learning rate parameter were used for the ANNs, and network learning at the different rates was compared to the human data. Second, the transfer of learning between profile analysis and intensity discrimination tasks was compared for the ANNs and humans. Again the pattern of performance was similar for the ANNs and human listeners. Given comparable stimuli, sensitivity was better for profile analysis than for intensity discrimination, and negative transfer occurred when switching from profile analysis to intensity discrimination or vice versa. These studies extend previous work [Howard et al., J. Acoust. Soc. Am. Suppl. 1 85, S37 (1989)] using feed-forward artificial neural networks trained with the back propagation learning rule to model human auditory phenomena. [Work supported by ONR.]