International Business Machines Corporation
Down-sampling for convolutional neural networks

Abstract:

A down-sampling operation for a convolutional neural network is provided. The present invention may include dividing received input to form a grid pattern having a set of odd and even numbered rows and columns and extracting data from each cell of the grid pattern to form a set of corresponding inputs. The present invention may include, in response to assigning data extracted from each cell to a corresponding channel, performing down-sampling of the extracted data for a respective channel associated with the input using an inverse of a first factor and increasing a number of the channels by a product of the first factor. The present invention may include adjusting a respective position of the received input by applying a transform to the data down-sampled to create transformed data. The present invention may include sending the transformed data as input to a next layer without addition or loss of data.