Using Artificial Neural Networks to Study the Evolution of Animal Communication
Animals often attend to only a few of the cues provided by the complex displays of conspecifics. We suggest that these perceptual biases are influenced by mechanisms of signal recognition inherited from antecedent species. We tested this hypothesis by manipulating the evolutionary history of artificial neural networks, observing how the resulting networks respond to many novel stimuli and comparing these responses to the behavior of females in phonotaxis experiments. Networks with different evolutionary histories proved equally capable of evolving to recognize the call of the tungara frog, Physalaemus pustulosus, but exhibited distinct responses to novel stimuli. History influenced the ability of networks to predict known responses of tungara frogs; network accuracy was determined by how closely the network history approximated the hypothesized history of the tungara frog. Our finding emphasize the influence of past selection pressures on current perceptual mechanisms, and demonstrate how neural network models can be used to address behavioral questions that are intractable through traditional methods.