A Rapid Two-Step Learning Algorithm for Spline Activation Function Neural Networks with the Application on Biped Gait Recognition

摘要

A fast two-stage learning algorithm is proposed to construct and optimize the weights of spline activation function neural networks (SAFNN). Feedforward network is firstly trained by back propagate (BP) algorithm, and then errors are applied to generate new neurons in hidden layers. A rapid dynamic updating algorithm is introduced to modify the new weights. Generalization capability and approximation precision are ensured by the two steps respectively. Simulation results on biped gaits demonstrate improvements in these two capabilities and of learning speed with comparison to traditional BP in SFANN and common NN.