METHOD OF TRAINING ARTIFICIAL NEURAL NETWORK Russian patent published in 2015 - IPC G06N3/08 

FIELD: physics.

SUBSTANCE: method comprises transmitting a multidimensional input vector X=[x₁, x₂, … x_N] to a multilayer neural network with the structure: N synapses in the input layer, one hidden layer with N_h synapses and N_c synapses in the output layer; calculating the corresponding cross-correlation and autocorrelation functions; determining the weight of the hidden and output layers based on a solution of a system of antigradient equations, wherein if the error increases, the direction of changing the weight of the hidden layer is selected through successive assignment of a pair of close values of the current weight of the hidden layer w₁(j,k)=w(j,k)-ξ and w₂(j,k)=w(j,k)+ξ, where 0<ξ≤0.001, and comparing the corresponding changes of the mean square error and , where E₁, E₂ are mean square error values corresponding to the weight of the hidden layer w₁(j,k) and w₂(j,k); is the mean square error value of the previous attempt, if ΔE₁<ΔE₂, the weight of the hidden layer is changed in the direction of reduction, if ΔE₁≥ΔE₂, the weight of the hidden layer is changed in the direction of increase.

EFFECT: ensuring convergence and acceleration of the process of training an artificial neural network.

5 dwg