Low-complexity convolutional neural network based on symbol random calculation

Abstract

The invention discloses a low-complexity convolutional neural network based on symbol random calculation. The system comprises a symbol random calculation module and an efficient convolution layer, apooling layer and a nonlinear excitation layer which are constructed based on the symbol random calculation module, the symbol random calculation module comprises a forward conversion module and a backward conversion module, the efficient convolution layer comprises a random processing unit and a random adder matrix, and the pooling layer is connected to the output end of the efficient convolutionlayer. According to the invention, a special sign bit is introduced; symbol random calculation is formed, the problem that the traditional random calculation can only represent positive numbers is solved; the calculation error caused by scaling coding is avoided, a novel parallel fast convolutional neural network architecture is provided based on the random calculation mode, the throughput rateof random convolution is increased while the hardware consumption is reduced, and the delay of random convolution calculation is reduced to a certain extent.

Description

technical field [0001] The invention belongs to the field of convolutional neural networks, in particular to a low-complexity convolutional neural network based on symbolic random calculation. Background technique [0002] Convolutional Neural Network (CNN) is the most important model in deep learning. In recent years, Convolutional Neural Network has been widely used in object detection, image classification, video surveillance and semantic segmentation. . Due to the high recognition accuracy of convolutional neural networks, it has been proved to be a powerful tool in the field of image recognition and sound recognition. In theory, as long as the number of layers of CNN is made deep enough, the network can obtain satisfactory recognition accuracy, and the CNN at this time becomes a deep convolutional neural network (Deep Convolutional Neural Network, DCNN). Since the operation of the deep convolutional neural network has the characteristics of intensive operation and sto...

AI Technical Summary

Problems solved by technology

[0004] Purpose of the invention: In order to solve the problem that the existing convolutional neural network hardware consumes so much that it is difficult to realize it on small wearable devices, it provides a low-complexity convolutional neural network architecture based on symbolic random computing, which can basically Under the premise of maintaining the same recognition accuracy of the neural network, the complexity of hardware implementation is effectively reduced

Images