Dendritic neuron model trained by information feedback-enhanced differential evolution algorithm for classification

Abstract

As the well-known McCulloch–Pitts neuron model has long been criticized to be oversimplified, different algebra to formulate a single neuron model has received increasing interests. The dendritic neuron model (DNM) which considers the nonlinear information processing capacity of both synapses and dendrites has shown its effectiveness on classification problems. However, an effective learning method for DNM is still highly desired and challenging because the traditional error back-propagation (BP) algorithm usually suffers from issues originated from the proliferation of saddle points and local minima trapping. In this study, a novel non-BP learning algorithm based on the differential evolution algorithm is proposed. By fully incorporating the feedback information from evolving population and individuals, the proposed algorithm significantly outperforms other ten state-of-the-art non-BP algorithms and BP in terms of the classification accuracy. Additionally, the neural morphology and hardware realization are also confirmed. These lead us to believe that the well-learned DNM is considerably more powerful on computations than the traditional McCulloch–Pitts type, and can be possibly used as a fundamental unit in the next-generation deep learning techniques.