In step S707, the parameters of the first determination unit 501 are updated. The update is performed by a gradient method. Assuming that the parameters of the first determination unit 501, i.e., the weight of the convolution layer 601 and the weight of the full connected layer 604, are
θ_d={w^(D1),w^(D4)}??[Math. 14]

the parameters are updated by the conversion below.
θ_d←θ_d+γ?_θde??[Math. 15]

Here, γ is a learning coefficient, and γ=0.01 in the present embodiment. Also,
?_θde??[Math. 16]

is a partial derivative of the error e with each parameter, and may be calculated by automatic differentiation.

In step S708, the parameters of the first estimation unit 203 are updated. The update is performed by a gradient method. Assuming that the parameters of the first estimation unit 203, i.e., the weight of the convolution layer 401 and the weight of the deconvolution layer 404, are
θ_g={w^(G1),w^(G4)}??[Math. 17]

the parameters are updated by the conversion below.
θ_g←θg+δ?_θge??[Math. 18]

Here, δ is a learning coefficient, and δ=0.01 in the present embodiment. Also,
?_θge??[Math. 19]

is a partial differential value of the error e with each parameter, and may be calculated by automatic differentiation. After the parameters of the first estimation unit 203 are updated in step S708, the process returns to step S703.