Ratio specificity, output reliability and code length.

<p>a) An example illustrating the bell-shaped correlation curve. Correlating the PN population response for a particular ratio, , against all other ratios gives this approximately bell-shaped curve, which peaks at with a correlation coefficient of value 1. The more specific the PN response to a ratio, the sharper and thinner the bell-shaped curve. Thus, the width of the bell-curve, , marked by the dotted lines (see text for definition) indicates the specificity of the PN response. b) The effect of response length on code specificity. This shows the specificity of the spatiotemporal PN responses as the response length taken into account is changed, indicating how the ability of the models to encode ratios is dependent on the temporal dynamics. The shaded areas show standard error. c) The effect of response length on reliability. Discriminant analysis was used to predict the input ratios from PN responses according a training set of responses (see text). The accuracy of this classification indicates the reliability of the PN responses for the identification of input ratios. This test was applied to different response lengths as in b). The accuracy of both types of networks becomes stable by 100 ms, with FPA networks at around 85% and LCA networks at around 91%. LCA networks start with lower accuracy but quickly increase and exceed that of FPA networks. The FPA networks' ability to hold ratio information actually decreases as they reach the stable attractors. Again, shaded areas represent the standard error.