INFORMATION PROCESSING UNIT, LEARNING NETWORK LEARNING VALUE CALCULATION PROGRAM AND LEARNING NETWORK LEARNING VALUE CALCULATION METHOD

摘要

PROBLEM TO BE SOLVED: To reduce a computational complexity of a convolution layer.SOLUTION: An information processing unit 100 has a Pooling layer 110b and a convolution layer 110a. The Pooling layer 110b acquires information on an error gradient having a plurality of elements from an upper layer in learning value calculation of a learning network having a plurality of layers. The convolution layer 110a when calculating a value of a first element included in a weight gradient specifies a region corresponding to the first element among a plurality of elements included in information acquired from a lower layer, and divides an element of the specified region into a plurality of partial regions. The convolution layer 110a executes processing of calculating the first element by calculating, for each partial region, a value based upon a total value of elements included in the partial region and a value of an element of the error gradient corresponding to the partial region, and adds respective calculated values.SELECTED DRAWING: Figure 4