METHOD AND DEVICE FOR LEARNING DEEP NEURAL NETWORK, AND DEVICE FOR LEARNING CATEGORY-INDEPENDENT SUB-NETWORK

摘要

The present invention provides a method for learning a DNN capable of reducing a time for learning the DNN using data belonging to a plurality of categories. This method includes a step for learning a language-independent sub-network (120) and language-dependent sub-networks (122 and 124) by using Japanese and English learning data. This step includes a first step for learning, using Japanese data, a DNN in which neurons in an output layer of the sub-network (120) and neurons in an input layer of the sub-network (122) are connected; a step for forming a DNN in which the sub-network (124) is connected to the sub-network (120) in place of the sub-network (122), and learning the DNN using English data; a step for alternatingly executing these steps until the learning data is finished; and a step for separating, after completion, the first sub-network (120) from the other sub-networks, and causing the first sub-network (120) to be stored in a storage medium as a category-independent sub-network.