Optimization method of dense-sparse-dense algorithm

Abstract

The invention proposes an optimization method of a dense-sparse-dense algorithm. The optimization method comprises an initial Dense training phase, a fuzzy phase, and a final Dense phase. An association degree of a network weight and a network is measured based on a membership degree; and an association degree of each piece of data information and a cluster is determined. The optimization method has the following advantages: firstly, compared with other classical networks, the provided optimized network is based on the value of a learning weight and network weights being important connection are calculated, so that the classification precision is improved based on the screening process; secondly, the frame mobility of the optimization method is improved by being compared with the conventional DSD and the optimization method can be applied to other novel networks like the VGG16 and vgg19 after Alexnet; and thirdly, in order to solve a problem of needing tens of thousands of iterative classification by the traditional deep neural network, the optimization method needs hundreds of iterations and thus the classification accuracy is improved effectively.

Description

technical field [0001] The invention belongs to the field of image classification and relates to an optimization method of a dense-sparse-dense algorithm. Background technique [0002] Deep learning belongs to the field of machine learning. With the introduction of more excellent neural networks, we know that the performance of complex networks better proves the highly nonlinear correlation between feature information and output. However, with the deepening of the network, the number of layers is increasing, and the network parameters are also greatly increasing, which leads to the increasing difficulty of network training. Song Han proposed a Dense Sparse Dense (DSD) algorithm, focusing on how to improve the accuracy of traditional models by improving the training process. We use sparsity to prune small, unimportant network weights, and retrain the network based on sparsity constraints to normalize the network. [0003] Although both DSD and dropout have pruning operation...

AI Technical Summary

Problems solved by technology

The DSD algorithm improves the performance of the network through selective network weights, but it cannot accurately distinguish those unimportant network weights

Images