An analytical solution of the compartmental model for use in local learning in artificial neural networks

Abstract

Most artificial neural network models do not take into account the extensive structure of the input (dendrite) of a neuron; the neuron is treated as a point at which all inputs meet. For neural network applications in which temporal relations have to be learned, extensive neurons could be interesting because they possess an intrinsic time dependent behavior. Learning in such a network may be done with local-learning models, in which weight updating is based on correlation between inputs and local dendritic parameters, such as the dendritic voltage. Extensive dendrites can be modeled with a compartmental model, a description in terms of electrical circuits. An analytical solution of the time course of the potential in the dendrite is described.