To provide a device capable of rasing generalization capability of a classifier in a learning device of pattern classifier by LGM-MCE learning.

An expression (1) is used to define an error classification scale value D_y(x;Λ) for measuring a degree of error classification of a sample x belonging to a class C_y. An expression, g_y(x;Λ) based on ψ>0, shows a discrimination function for a belonging degree of x to C_y. A learning device executes: a step for obtaining the error classification scale values of the sample belonging to C_y; a step for, with setting a real probability distribution generating them as a Parzen distribution of a window width h_y centering on each of the error classification scale values, evaluating likelihood of the distribution with a function of h_y; a step for estimating the likelihood through cross-validation most likelihood estimation; a step for using an expression, α_y=4/((2π)^1/2*h_y), to calculate an optimal value α_y of lost smoothness, with respect to h_y that provides a most likelihood distribution; and a step for adjusting a learning parameter Λ so as to minimize an empirical average loss that is a function of α_y.