Dynamic neural field optimization using the Unscented Kalman Filter

In this article, we proposed to use Unscented Kalman Filters in order to tune the parameters of a dynamic neural field. To do so, a scenario involving the input given to the neural field as well as the desired output must be defined. Belows are examples of Input/desired output provided to the algorithm as well the output obtained after the algorithm converged and a video showing the neural field during convergence of the algorithm. The illustrations are all in space x time for 1D neural fields.

Competition

In the competition scenario, two stimuli with slightly different amplitudes are exciting the neural field. The problem is to find the lateral connections establishing a competition between the two excited regions and leading to the selection of the most salient targets.

Input Target Output
input_competition target_competition output_competition

Below, we show the neural field firing rates during the convergence of the algorithm.

Working memory

In the working memory scenario, two stimuli of initially weak amplitude are exciting the neural field. Then, the amplitude of the stimuli are increased and decreased back to the initial level. The sequential increase of the targets’ amplitude may result from a sequential selection of the targets that must be memorized for some reasons. As an output, we want the neural field to hold the position of these stimuli as soon as the amplitude reaches some threshold.

Input Target Output
input_wm target_wm output_wm

Below, we show the neural field firing rates during the convergence of the algorithm.