Optimization of DBN Network Structure Based on Information Entropy
Ziliang Huang, Yan Long Cao, Tianbao Wang · Journal of Physics Conference Series · 2019
To determine the appropriate depth of DBN network and hidden layer neurons at the same time, the information entropy of the input layer is analyzed on the basis of information entropy and the traditional reconstruction error calculation and the decision of network depth. Thus, according to the relationship between information entropy and hidden layers, an optimization method based on information entropy to determine the number of hidden neurons is proposed, which makes the structure of the DBN network model tend to be better. Experimental results on the handwritten numeral recognition demonstrated that the proposed method is capable of self-organizing the depth of the network and hide the number of neurons in the hidden layer, effectively optimize the DBN network structure, reduce the training time of the network, as well as improve the network accuracy and recognition accuracy.