To actively recognize an object by allowing the intermediate layer of a sandglass neural network to learn by extending stepwise so as to determine the intermediate layer of a minimum dimensional number and so as to estimate the state of the object by the use of an object estimating part learning by the extended learning method.
The sandglass neural network N1 (sandglass N1) to which the intermediate layer of a proper size is set is made to learn. After then the part of the intermediate layer is set to be for input and the intermediate layer constitutes a sandglass N2 set to be a value smaller than last time to learn. Then the operation is repeated every satisfactorily learning and when learning is no longer converged, learned sandglasses are connected to be one to obtain a sandglass neural network where the number of intermediate layers corresponds to a minimum dimension held by input information. This method is applied to an object state estimating part at an object actively estimating device to efficiently construct an object state estimating part even when an essential dimension of input information is not known.
YAMAKAWA HIROSHI
KIMOTO TAKASHI
Next Patent: COUPLING COEFFICIENT UPDATING METHOD IN PULSE DENSITY TYPE SIGNAL PROCESSING NETWORK