PURPOSE: To obtain stable operation precision even for unlearnt data.
CONSTITUTION: Input compressed data is generated by learning the input data of an object problem by an input data compressing neural net 10, and output compressed data is generated by learning the output data of the object problem by an output data compressing neural net 30, and further, mapping from the input compressed data to the output compressed data or the mapping from the output compressed data to the input compressed data is learnt by an input/ output compressed data mapping neural net 20. Then, the input data of the object problem is inputted to the input layer 101 of the neural net 10, and the output of the output layer 112 of the neural net 30 is defined as the estimated value of the output of the object problem, or the output data of the object problem is inputted to the input layer 108 of the neural net 30, and the output of the output layer 105 of the neural net 10 is defined as the input data estimated value of the object problem. Thus, stabler approximation precision can be obtained for the unlearnt data compared with a case that the three-layer neural net of a conventional method is used.
HAMAGUCHI YUKIO
KANAZAWA NOBUHIRO
Next Patent: NEURAL NETWORK