Device and method for convolution operation

Abstract

A convolution operation method according to one embodiment of the present invention can comprise the steps of: extracting, such that each are different, a plurality of partial matrices, which are matrices that correspond to partial areas of feature data, which is an MxN matrix (M and N are natural numbers) and have the same dimension as a convolution kernel, which is a KxL matrix (K is a natural number of M or less, and L is a natural number of N or less); generating a first vector including, as an element, an element belonging to the at least one partial matrix from among elements of the feature data, and a second vector including, as an element, an element of the convolution kernel; extracting, for each of the partial matrices, a partial vector including all elements thereof and being apart of the first vector; and multiplying, for each of the partial matrices, respective elements of the corresponding partial vector and elements present at the corresponding position in the second vector, and adding all the results of the performed multiplication so as to calculate a convolution operation result with the convolution kernel.

Description

technical field [0001] The present invention relates to an apparatus and method for efficiently performing convolution operations when recognizing target data using a specific extraction model such as a convolutional neural network (CNN). [0002] For reference, this application claims priority based on Korean Patent Application (Application No. 10-2018-0035104) filed on March 27, 2018. The entire contents of such corresponding applications which are the basis of priority are incorporated in this application as references. Background technique [0003] Recently, a technique of automatically analyzing recognition target data such as image data or text data and classifying the above recognition target data by type or extracting information contained in the above recognition target data has been adopted as a backup in artificial intelligence (AI) technology. By the attention. Such techniques can be applied in various fields such as autonomous driving, medical image reading, s...

AI Technical Summary

Problems solved by technology

Such repeated convolution operations with a large amount of calculations may become an obstacle to data recognition using convolutional neural network techniques.

In particular, in the field of image data recognition where fast calculation is important, such as the above-mentioned autonomous driving, the time required for convolution calculation may become a big problem

[0006] Also, although there have been many studies on methods for efficiently performing the above-mentioned convolution operations, even through these studies, there are many cases where there are problems in terms of device implementation, such as the method for achieving high efficiency in terms of algorithms. Complicated hardware configuration, etc.

Images