BP Neural Network Transparency and Structure Reduction Algorithm Based on Weight Contribution Rate

Neural networks have excellent nonlinear mapping approximation ability, but the neural network modeling method belongs to the “black box” method. The obtained model lacks transparency and the interpretability of each variable is poor. In this paper, the single layer neural network model is transparently studied, combining the neural network paraphrasing map, the connection weight method and the improved randomization test, this method can be further extended to provide a reference method for more complex simplification of deep or multi-layer network models. Through the research on the numerical simulation of neural network and the classification of double crescent data, the results show that the simplified method obtains the internal information of process variables and greatly improves the model's “understandable” ability. Therefore, this study provides a good way for the transparency of the neural network model and the streamlining of the network structure in deep learning.