PUB - Publikationen an der Universität Bielefeld

Pre-touch sensing has a larger range than tactile sensing, but shorter than vision. By measuring mutual capacitance, electric field pre-touch sensors can detect local objects if their dielectric properties differ from that of air. When the actively emitted electric field by such a device is measured at multiple positions, spatial modulation patterns can provide geometric information on the perturbing objects. To illustrate this, an electric field pre-touch sensor was used to make recordings of simple objects such as cubes and spheres. Interpreting the arising sensor data can however be a substantial challenge, as the signals are multi-parameter dependent while the carrier wave itself is subject to noise. Neural structures associated with the electric sense of mormyrid weakly electric fish are evolutionary optimized to cope which such problems and hence provide a convenient biological model in which processing of electrosensory information can be studied. Using a pre-existing model constructed from available electrophysiological and neuroanatomical data, a spiking neural network using a temporal coding scheme was here developed for potential implementation in dedicated neuromorphic hardware. The technical perspective is incorporated at the neuronal level of the network, where neuron and synapse models reflect the idealized behavior of analogous implementations in neuromorphic hardware. The here implemented network extends its predecessor by a second perceptual channel, providing it with the capability to selectivity detect signal in- or decreases. Based on simulated inputs showing the perturbations caused by a sphere in a uniform electric field, the networks ability to provide indication on presence of objects encoded in the precise timing of spikes was investigated and its spatial-temporal response further evaluated via information theoretical measures. Finally, the principle feasibility of the approach was tested by a step-wise emulation of one of the two perceptual channels in neuromorphic hardware. The results illustrate that an implementation of the network in a dedicated chip is indeed possible and that the network – assuming that the required sensor data can be provided – is capable of conveying pre-touch information by a neuronal output that is filtered, amplified, sparsified, and overall more informative than the initial input.