Figure 2.

Example traces for the model for Gryllus bimaculatus. The stimulus (A) had a pulse period of 40 ms (chirp duration 200 ms, chirp period 500 ms). The first filter (B, green) exhibited a pulse period of 40 ms and responded well to the pattern (C, green). The second filter's dominant modulation was relatively slow; accordingly, it responded poorly to this stimulus (C, red). For the computation of the filtered stimulus the filter (B) is first aligned with the beginning of the stimulus (A) and then multiplied with the amplitude values of the stimulus over the duration of the filter. The result is a product as a single point in time that reflects the similarity of the filter with the stimulus, high values indicating high similarity. The filter is then shifted by one step (given by the time resolution of the stimulus) and then multiplied with the respective amplitude values of the stimulus as before. This procedure is repeated until the end of the stimulus is reached and the filtered stimulus emerges as a new time series. Patterns were normalized such that the distribution of amplitudes over the whole stimulus set exhibited zero mean and unit standard deviation. The nonlinearity for each filter was relatively steep (D), transforming the stimulus into an almost binary trace (E). This trace was then integrated over time and the resulting values for each LN model were weighted to obtain a prediction of the phonotaxis score (modified from Clemens and Hennig, 2013).