Abstract
We implemented and investigated a method for the measurement of frequency-domain Volterra- Wiener kernels of nonlinear neurons. The procedure is based on higher-order spectra. For the 2nd- order kernel the cross-bispectrum is used. The method is not dependent on white noise stimulation, i.e. the input frequency content can be adjusted in order to drive the neurons. Since the method works in the frequency domain it seems to be less dependent on the disadvantageous convergence properties of spatial/temporal averages of classical white noise procedures. Theoretically, it should also be insensitive to additive Gaussian noise. We tested the method for a simulated nonlinear V1 cell that is intended to model nonlinear interactions with the extra-classical receptive field. The test is performed with additive and multiplicative noise, and with different noise levels. Since the true kernel of the simulated neuron is known, it is possible to evaluate the quality of the estimated kernel in dependence on the noise level. For additive noise, the method shows good performance up to a S/N of −20dB.
Supported by DFG (SFB 462 and GRK 267)