Research on the Generalization Problem of BP Neural Network
Yanli Tian,Deyin Fu,Guangzheng Li
DOI: https://doi.org/10.1109/access.2024.3452710
IF: 3.9
2024-01-01
IEEE Access
Abstract:The application of neural network models is becoming increasingly widespread.One of the most relevant aspects of neural networks is their generalization ability, which predicts situations that are not included in the training set. Research has shown that training set samples play a dominant role in the learning and training of back propagation (BP) neural network model, and the information contained in them directly affects the network performance.Alarge number of similar samples not only prolongs the training time of the model, but also leads to problems such as decreased network generalization ability. Based on this, this study focused on the impact of training set samples on the generalization ability of the BP neural network model. A model was constructed on this basis by introducing dynamic clustering methods to screen the training set samples with representative and typical features. The empirical results indicate that using dynamic clustering methods to screen samples can effectively remove redundant information from the training set, enhance the network generalization ability, and improve the model prediction accuracy.Moreover, for large datasets, it can effectively improve the model’s fitting degree and convergence speed, and enhance the "noise tolerance" performance of the BP neural network model.