PURPOSE: To perform recognition learning efficiently at a high speed.
CONSTITUTION: Various series of feature data of an object are inputted to unit recognition units in respective layers of the tree branch structure of plural unit recognition units (k11-k12, k21-k30, p31-p350, and q41-q43) and each of the unit recognition units in the respective layers has a quantizer which performs quantization corresponding to an input signal; and coupling paths to unit recognition units in lower layers of the network are switched and a selection path up to the lowest layer is determined. In forcible learning mode, a learning unit increases the coupling strength between path input terminal of unit recognition units (p31-p350) in the layer right above the lowest layer and a path output terminal that the output value of the quantizer indicates and in fine adjustment learning mode, the learning unit decreases the coupling strength between the path input terminal of the unit recognition units (p31-p350) in the layer right above the lowest layer and a path output terminal that the output value of a maximum output unit detector 6 indicates only when a wrong recognition result is obtained.
MARUNO SUSUMU
KOJIMA YOSHIHIRO
IMAGAWA TARO
SHIMEKI TAIJI
