PURPOSE: To shorten learning time to be required for the convergency of a weight value by learning with using short time-sequential data which is a part of long time-sequential data at first in the case of learning the long time- sequential data.
CONSTITUTION: The 1st stage learning section is set up as the initialization of learning sections. Time-sequential data in the section are learned. When the learning of the 1st section is converged, the learning section is extended and data in the 2nd stage learning section are learned. While extending the section to the 3rd and 4th stage learning sections, learning is similarly advanced. Since learning sections are gradually extended from the shortest learning section, stable and high speed learning can be attained as compared with a case of learning a long learning section at first.
ABE MASAHIRO
Next Patent: LEARNING METHOD