Embedded low power consumption convolutional neural network method

Abstract

The invention discloses an embedded low-power convolution neural network method; for the two adjacent layers of the conv layer and the conv+concat layer, the idea of ​​inception in GoogLeNet is adopted, and the volume of 1×1 and 3×3 The product kernels are arranged in parallel, and features on different scales are collected separately, and then the feature maps calculated by the two convolution kernels are spliced. For the conv+concat layer, the idea of ​​​​ResNet is adopted. The splicing results of this network in the conv+concat layer The element-by-element addition operation of short cut connection was performed, and each layer was pre-pruned separately. The cut neurons accounted for 50% of the total, and the impact of each layer on the overall accuracy was tested. Reduce the convolution size of each layer of the multi-layer network, reduce the number of convolution kernels and the network has a certain degree of sparsity, so the number of multiplications required for each layer is very small. In specific calculations, it can be converted to have CSR The product of the sparse matrix and the dense matrix in the storage format, in order to achieve the purpose of small memory usage, fast calculation speed and high precision.

Description

technical field [0001] The invention relates to the field of pattern classification, in particular to an embedded low-power convolution neural network method. Background technique [0002] Convolutional Neural Networks (CNN) is a deep learning architecture that has been developed rapidly since 2012 and has attracted widespread attention. This architecture effectively improves the accuracy of image classification and object recognition. Compared with the traditional manual feature extraction method, CNN not only has high recognition accuracy, but also avoids the complicated preprocessing process; compared with the traditional Back Propagation (BP) neural network, CNN uses the sharing strategy , the number of neurons is greatly reduced, thus avoiding the phenomenon of over-fitting to a certain extent. [0003] With the development of GPGPU, CNN has more and more network layers, from the 8-layer network of AlexNet, the winner of the ImageNet Image Classification Challenge in 2...

AI Technical Summary

Problems solved by technology

However, these network structures require a powerful GPU or even a GPU cluster to train them, and the memory / display memory consumed by neurons is huge, so the existing CNN network is not suitable for direct deployment on embedded devices with low power consumption and resource constraints

Images