The present invention comprises: a first step for calculating, by a deep learning network, STKT that is an output value obtained by applying a softmax function to a first soft label that is a distribution of result values output by receiving an input of data to be learned; a second step for calculating, by a lightweight deep learning network, Pmodel that is an output value obtained by applying the softmax function to a second soft label that is a distribution of result values output by receiving an input of data to be learned; a third step for calculating a similarity (KDR) between STKT and Pmodel; a fourth step for calculating a ratio (TSTR) of a current learning time step to an entire learning time; a fifth step for selecting a variable (TSSR) for transference ratio determination, on the basis of the similarity (KDR) and the ratio (TSTR) of the current learning time step; and a sixth step for adjusting, by the lightweight deep learning network,. a ratio of an actual correct answer (EKT) and a ratio of the STKT transferred from the deep learning network, on the basis of the variable (TSSR) for transference ratio determination.
展开▼