Context sensitivity in profile analysis by parallel distributed networks

Abstract

Spectral shape discrimination with uncertain signal frequency was simulated using simple, three‐layer, feed‐forward, parallel networks. As in previous studies, pairs of multitonal sounds were used—one with all components at a uniform, background level and the other with a tonal signal added in‐phase to a single component [Howard et al., J. Acoust. Soc. Am. Suppl. 1 85, S37 (1989)]. Networks were trained to select the signal alternative in a 2 AFC paradigm using the generalized delta rule with error propagation and were then tested under various conditions. The signal frequency was either fixed or varied across trials, and stimulus pairs were presented either alone or with a cue that was correlated with the signal frequency. Network sensitivity was greater with the fixed signal frequency than with the varied signal frequency and this difference was reduced by the presence of a predictive cue. As in earlier experiments, these stimulation results mimic previous findings with human listeners under similar task conditions. [Work suported by ONR.]