Альтернативный подход к организации обучения сигмоидальных нейронов в архитектуре многослойной нейронной сети

Abstract

The article formulates and mathematically substantiates a modification of the Back Propagation Error (BPE) algorithm for accurate training of neural elements with a sigmoid activation function in the architecture of a multilayer neural network. The proposed training rules for modifying the synaptic connections of neurons in the output and hidden layers can be used to construct efficient algorithms that reduce the time and computational complexity of the learning process for multilayer architectures.