Biomimetic spike representation for the encoding of sonar signals

Abstract

Biosonar echoes received by bats in their natural habitats are short, highly time-variant acoustic waveforms. Because of these signals properties, the way in which the bats' auditory nerve represents the echoes could be a model for how sparse neural time-code can capture salient signal features. Here, the spike-encoding of natural echoes has been studied based on a large data set containing about 220 00 echoes that were collected by hand-carrying a biomimetic robotic sonar head through forest environments. The sonar head was equipped with flexible noseleaf and pinna shapes that could deform during pulse emission/echo reception in a similar fashion to what horseshoe bats do. This peripheral dynamics was turned on for half of the recoded echoes and turned off for the other half. The echo waveforms were transformed into spike trains using two spike-generation models, each with different levels of complexity: The simplest version was a “leaky integrate-and-fire model"; in the more complex version, a response kernels was added to model the refractory behavior of the neurons. The input to the spike generation models were the outputs of three different basiliar membrane models of varying complexity levels. The coding capacity of the spike trains has been evaluated using information-theoretic methods.