PROCESSING METHOD BY CONVOLUTIONAL NEURAL NETWORK, LEARNING METHOD OF CONVOLUTIONAL NEURAL NETWORK, AND PROCESSING APPARATUS INCLUDING CONVOLUTIONAL NEURAL NETWORK

摘要

PROBLEM TO BE SOLVED: To reduce power consumption and an operation amount of a convolutional operation of a convolutional neural network.SOLUTION: Matrix data for a convolutional operation is divided into two, a first half part and a second half part, with reference to a threshold value. The first half part contains a relatively large number of major terms, while the second half part contains a relatively small number of major terms. A convolutional operation unit performs a convolutional operation of the first half part and a convolutional operation of the second half part by dividing into two. The convolutional operation of the first half part executes an operation for generating first operational data used for a maximum value sampling operation of a pooling operation unit. The pooling operation unit selects vector data to which a convolutional operation of a matrix vector product should be applied in the convolutional operation of the second half part. The convolutional operation of the second half part performs a convolutional operation with respect to the selected vector data to generate second operational data. Intermediate layer data of a convolutional neural network is obtained by adding the result of the maximum value sampling operation and the second operational data.SELECTED DRAWING: Figure 4