Convolutional neural network system, memristor array and convolutional neural network

Abstract

The embodiment of the invention provides a convolutional neural network system, a memristor array and a convolutional neural network. The system carries out the convolution operation of a current image block in a to-be-processed image through a plurality of columns of bit lines and a plurality of rows of word lines in the memristor array by employing a convolution kernel with the size of N * N, comprising the following steps: successively selecting two overlapped and adjacent image blocks on a to-be-processed image according to the step length of a convolution kernel, and determining feature values of a plurality of pixel points multiplexed between the two image blocks; according to the characteristic values of the plurality of multiplexed pixel points, determining a plurality of multiplexed weights in the first N columns of bit lines; setting the working state and conductance of the memristor unit according to the plurality of weights; and obtaining current values output by the first N columns of bit lines in the current convolution operation period, and taking the current values as convolution results. According to the convolutional neural network system provided by the embodiment of the invention, convolution operation is performed by using the feature values of the plurality of pixel points multiplexed between the image blocks, so that the calculation amount of the memristor array in each period is increased, and the calculation efficiency is improved.

Description

technical field [0001] The invention relates to the field of semiconductor devices, in particular to a convolutional neural network system, a memristor array and a convolutional neural network. Background technique [0002] The non-volatile memory characteristics of memristors enable them to have data storage capabilities, and the ability to adjust their conductance states enables them to have numerical computing capabilities. The combination of the two enables memristors to achieve the fusion of computing and storage at the same location. The realization of "integration of storage and calculation" is expected to break through the limitations of the traditional von Neumann computing architecture, and has broad application prospects. [0003] The related technology proposes to use a memristor array composed of memristor units and metal wires to realize vector and matrix multiplication. By setting the conductance and on-off of the memristor units, the corresponding current val...

AI Technical Summary

Problems solved by technology

[0004] However, restricted by the structural characteristics of the memristor array and the principle of neural network mapping calculation, when the calculation process of the neural network is mapped to the memristor array, the utilization rate of the memristor array is not high. The convolution operation cycle can only complete a limited amount of calculation, resulting in low computational efficiency of the memristor array

Images