The use of neural networks for the determination of the signal modulation frequency from the firing activity pattern in the auditory neurons of the frog

A two-layer backward propagation neural network was used for the determination of the modulation frequency of tonal signals from the firing patterns of single neurons located in the cochlear nuclei and the torus semicircularis of the grass frog (Rana t. temporaria). As an input of the neural network, a sum of several single responses of a neuron to an amplitude-modulated stimulus was used. The number of inputs corresponded to the number of the time readouts of the summarized response (usually 60), and the number of output elements corresponded to the number of modulation frequencies to be distinguished (from 3 to 15). In the case of a good synchronization of the input firing activity with the signal envelope, the classification was successful even if the training and classification were performed with the use of individual responses. An increase in the number of summed responses to 10–20 lead to a simplification of the training procedure. The results of the study were discussed in the context of the problem of the formation of periodicity detectors at the upper levels of the auditory pathway in vertebrates.