Biologically Plausible Neuron Model with a Mutate-archive Differential Evolution Algorithm

The dendritic neuron model (DNM), due to its exceptional nonlinear information processing capabilities and its development based on the structure of human brain neurons, is considered a promising alternative to the McCulloch-Pitts neuron model. However, traditional backpropagation (BP) learning algorithms are often hindered by issues such as gradient vanishing, gradient explosion, excessive reliance on gradient information, low training efficiency, and limited flexibility. Therefore, there is an urgent need for an efficient learning algorithm. This study proposes a differential evolution algorithm, where the mutation factor is adaptively updated through an external archive (MaDE), to replace traditional BP. The proposed algorithm samples the mutation factor from a Cauchy distribution with an activation function and updates the distribution using historical information. Experimental results demonstrate that, across a wide range of classification datasets, our algorithm significantly outperforms other algorithms, including BP.