A general spike coding model for natural echoes

Abstract

Bats navigating in natural environments face a challenging task in identifying landmarks (e.g., vegetation) that are composed of many reflecting facets and hence generate random echo waveforms. The biological substrate for these tasks is a spike code. Informative code features were identified by a computational approach which encoded a natural stimulus ensemble (84<th>800 echoes from four deciduous tree species) using a parsimonious model of spike generation. Interspike intervals spanning more than one cycle at the auditory bandpass filter’s center frequency proved convenient classification substrates because of the following properties: (a) almost certain presence in each echo representation; (b) simple probabilistic description (uncorrelated); (c) undemanding constraints on firing threshold locations and spike time resolution. An m-ary sequential probability ratio test using a first-order characterization of these intervals (number plus sample mean of duration and threshold location) demonstrated good classification performance (0.1% to 1% error probability at <10 expected echoes) as well as remarkable robustness against changes in the spike-model parameters. Distant-cycle intervals are therefore hypothesized to be a primal sketch building block analogous to visual edges, with the notable difference that they do not delineate deterministic shapes [Work supported by DFG, SFB 550 B6.]