METHOD OF PROCESSING INFORMATION IN NEURAL NETWORKS Russian patent published in 2010 - IPC G06F1/00 

FIELD: information technology.

SUBSTANCE: at the logical inference step, the digital code of each coordinate of the intermediate vector is converted to two digital code values through linear functional transformation performed with two complementary compression functions φ(x) and 1-φ(x), where x is the digital code of the coordinate of the intermediate vector, and digital codes of the first and second intermediate vectors are formed. At the composition step, digital codes of coordinates of the first intermediate vector are scaled through a prod operation in weighted connections and digital codes of weighted values are transmitted in accordance with the system of rules of logical inference describing the neural network to inputs of logical formal neurons performing a max operation over digital codes of weighted values, and digital codes of coordinates of a third intermediate vector are formed, and digital codes of coordinates of the second intermediate vector are scaled through a prod operation in weighted connections and digital codes of weighted values are transmitted to inputs of extra logical formal neurons performing a min operation over digital codes of weighted values, and digital codes of coordinates of a fourth intermediate vector are formed. Digital codes of coordinates of the third and fourth intermediate vectors are then scaled through a prod operation in weighted connections and digital codes of weighted values are transmitted to inputs of output logical formal neurons performing a max operation over digital codes of weighted values, and digital codes of coordinates of the output vector of the neural network are formed.

EFFECT: higher functional stability and information security in neuro-fuzzy networks owing to increase in information redundancy.

1 dwg