Plural neural network system having a successive approximation learning method

Claim 2

A neural network structure comprising: input unit means for receiving input data; a plurality of neural networks connected in parallel and connected to said input unit means, said plurality of neural networks successively learning correspondence between said input data and teacher data, so that a difference between a first output of said neural network structure derived from said input data and a second output of said neural network structure derived from said teacher data is minimized; and output unit means, connected to each of said plurality of neural networks, for outputting a result of learning on the basis of results of learning in said plurality of neural networks, a sum of said results of learning being a result of learning of said neural network structure.

Claim 3

A neural network structure as claimed in claim 1, wherein each of said output unit means include means for summing data supplied thereto, and the n-th neural network (n=1, 2, . . . ) learns said correspondence until a sum of the results of learning supplied form the up to n-th neural networks becomes equal to a predetermined result, while the learning of the up to(n-1)th neural networks is stopped.

Claim 4

A neural network structure as claimed in claim 2, wherein when said sum of the results of learning supplied from the up to n-th neural networks becomes sufficiently close to said teacher data, the learning is stopped and is not carried out in the (n+1)-th and following neural networks.

Claim 5

A neural network structure as claimed in claim 2, further comprising means for calculating an error, wherein the n-th neural network learns said correspondence until a local minimum is obtained in a square of an error calculated between an output of said nth neural network and an output based upon teacher data used to train said nth neural network.

Claim 6

A neural network structure as claimed in claim 2, further comprising means for counting, wherein the n-th neural network learns said correspondence until the number of times that the learning in the n-th neural network is repeatedly carried out becomes equal to a predetermined number.

Claim 7

A neural network structure as claimed in claim 1, wherein each of said plurality of neural networks includes a plurality of units each summing data supplied thereto and modifying said sum, and a plurality of links connected to said units, each of said links having a predetermined weighting factor.

Claim 8

A neural network structure as claimed in claim 6, wherein said plurality of units are of a multilayer structure and units for different layers are mutually connected through said links.

Claim 9

A neural network structure comprising: input unit means for receiving input data; a plurality of neural networks connected in parallel and connected to said input unit means, said plurality of neural networks successively learning correspondence between said input data and teacher data, so that the difference between an output derived from said input data and an output derived from said teacher data is minimized; and output unit means, connected to each of said plurality of neural networks, for outputting results of learning of said plurality of neural networks; wherein each of said neural networks includes: input units connected to said input unit means, each of said input units passing said input data as it is; intermediate units connected to said input units through links, each of said intermediate units summing said input data supplied thereto, modifying said sum and outputting modified data; and output units connected to said output unit means, each of said intermediate units being connected to one of said output units through a link, each of said output units modifying said modified data supplied thereto from a related of said intermediate units.

Claim 10

A neural network structure comprising: input unit means for receiving input data; a plurality of neural networks connected in parallel and connected to said input unit means, said plurality of neural networks successively learning correspondence between said input data and teacher data, so that the difference between an output derived from said input data and an output derived from said teacher data is minimized; and output unit means, connected to each of said plurality of neural networks, for outputting results of learning of said plurality of neural networks; wherein each of said neural networks includes: input units connected to said input unit means, each of said input units passing said input data as it is; intermediate units connected to said input units through links, each of said intermediate units summing said input data supplied thereto, modifying said sum and outputting modified data; and output units connected to said output unit means, each of said intermediate units being connected to one of said output units through a link, each of said output units modifying said modified data supplied thereto from a related of said intermediate units.

Claim 11

A neural network structure as claimed in claim 8, wherein each of said neural networks learns said correspondence on the basis of a back propagation process so that the following formulas as satisfied ΔW.sub.ji (n+1)=ηδ.sub.j 0.sub.i +αΔW.sub.ji (n) Δθ.sub.j (n+1)=ηδ.sub.j +αΔθ.sub.j (n) where: Wji is the weighting factor of the link connecting the unit i and the unit j, θ j is a bias of the unit j, n; the number of times that the learning is carried out, ηα; are constants, 0 i ; output derived from unit i (i=1, 2, . . . ) ##EQU2## F(x); a function which assumes values equal to or larger than 0.