Profile analysis by parallel distributed networks

Abstract

Several profile analysis (spectral shape discrimination) phenomena were simulated using simple, three‐layer, feed‐forward, parallel networks. As in previous studies with human listeners, 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 the central component. Networks were trained to select the signal alternative in a 2AFC paradigm using the generalized delta rule with error propagation and were then tested under various conditions. Background levels were varied randomly either between or within trials. In the former case, the network learned to compare energy at the signal frequency across the two sounds (intensity discrimination), whereas, in the latter case, it learned to compare the distribution of spectral energy within sounds (profile analysis). These results mimic previous findings with human listeners reported by Green and his colleagues [D. M. Green, Profile Analysis (Oxford U. P., New York, 1988)]. In follow‐up simulations, improved profile analysis have been demonstrated when the frequency range of tonal components is expanded and as the duration of the complex increases to 100 ms. [Work supported by ONR.]