Deep Learning Training Based On Clustering Algorithm Is Improved

Deep learning as a new research direction in machine learning in recent years,in various fields have made great achievements.Deep learning complex network structure,network training requires a large amount of data.However,in real life,it is difficult to obtain such massive data;at the same time because of deep learning many layers,even using the backpropagation training network,will consume a lot of time.So how to effectively train the network for a small amount of data has become an important problem to be solved in the deep learning.Based on the characteristics of deep learning and the traditional clustering algorithm based on the analysis and summary of training methods on deep learning effectively improved.The main work of this paper are summarized as follows:1.According to the shortcomings of the existing clustering algorithms,a clustering algorithm of data migration based on superposition entropy.The algorithm constructs the entropy energy function,determine the energy radiation field,so as to obtain the sample relative scope,the construction of information entropy in the sample space,and finally through the data migration for clustering samples.The experimental results show that this algorithm greatly improves the adaptive clustering algorithm.2.It is proved that the matrix clustering algorithm k-means decomposition,so that the characteristics of the learning algorithms of expression.Through the mathematical meaning and physical meaning of convolution analysis,obtained the relationship between the expression of feature clustering algorithm learning and convolution kernel parameters,and then draw the initialization method of deep learning parameters in the process of convolution convolution.3.Through the traditional clustering algorithm for deep learning network convolution kernel parameter initialization,and improve the deep learning training structure,with the characteristics of the traditional clustering algorithm learning results aided deep learning training,finally realizes the fast training of small sample deep learning network.Finally,the experimental results show that the DDSM medical image data set,when the sample is 50*50*50000,the proposed clustering algorithm based on the depth of learning and training method,in the premise of accurate rate is 97% basically unchanged,will shorten the time of 15%,the training data is reduced by 40%.In order to achieve the training depth of network,and effectively shorten the training time.