LEARNING METHOD OF DEEP NEURAL NETWORK, STORAGE MEDIUM STORING PARAMETER OF SUBNETWORK OF DEEP NEURAL NETWORK, AND COMPUTER PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide a learning method of a DNN that can shorten a time for learning the DNN by data belonging to a plurality of categories.SOLUTION: This method includes the steps of learning a language-independent subnetwork 120 and language-dependent subnetworks 122 and 124 by Japanese-to-English learning data. The steps include: a first step of learning a DNN connecting a neuron of an output layer of the subnetwork 120 with a neuron of an input layer of the subnetwork 122 by Japanese data; a step of forming the DNN connecting the subnetwork 124 to the subnetwork 120 instead of the subnetwork 122 and learning it by English data; a step of alternately executing these steps until the learning data is completed; and a step of separating the first subnetwork 120 from other subnetworks after completion, and making a storage medium store it as a category-independent subnetwork.