Evaluating the importance of different temporal features of calling songs in cricket phonotaxis

Abstract

Event Abstract Back to Event Evaluating the importance of different temporal features of calling songs in cricket phonotaxis Gundula Meckenhäuser1, 2*, Matthias R. Hennig3 and Martin P. Nawrot1, 2 1 Freie Universität Berlin, Institute of Biology, Germany 2 Bernstein Center for Computational Neuroscience Berlin, Germany 3 Humboldt-Universität zu Berlin, Department of Biology, Germany Acoustic signals are used by Gryllus bimaculatus to communicate and they play a key role in mate choice. Males produce calling songs to attract females, females in turn use these songs to discriminate conspecific songs from the signals of other species. If they are willing to pair females approach the singing male - a behavior referred to as phonotaxis. So far, extensive behavioral experiments in which phonotactic responses of female crickets were tested by varying song features, revealed that song patterns are processed in the time rather than in the spectral domain [1]. Here, we explored the relevant cues for song pattern discrimination on several time scales. We present feed-forward artificial neural networks (ANNs) that quantitatively predict the phonotactic value for untested calling songs that are described by temporal features such as duration, pause, period and duty cycle for both pulses and chirps. We employ ANNs with the following architecture: Input neurons representing temporal song features project to nonlinear neurons in the hidden layer which project to the output neuron that relates the phonotactic value. For training and testing the networks we used 218 artificial songs for which the phonotactic values had already been determined in experiments. Our complete ANNs that use eight temporal song features show a mean squared error (MSE) of MSE = 0.034. By a greedy backward elimination of features, we came up with minimal ANNs using just pulse period, chirp duration and chirp duty cycle as features. The dimensionality reduction of the feature space also improved the performance: The mean squared error for minimal networks is MSE = 0.019. Thus, our findings suggest that among all tested temporal features, pulse period on the short time scale and chirp duration and chirp duty cycle on the long time scale are the most relevant ones for the evaluation of conspecific signals. Further, the minimal ANNs show high predictive power and can be used to complement experimental testing of female phonotaxis in the laboratory. Acknowledgements This work is funded by the German Research Council (DFG) within the Collaborative Research Center "Theoretical Biology" (SFB 618).