An optimized neural network approach and its application to classification problems

Jie Han, Hongru Song, Buyun Chen

Abstract: Traditional neural network models are plagued with slow training speed and low classification accuracy, posing numerous challenges to the development of image classification technology. To address these issues, this paper employs a multi-strategy integrated optimization framework to optimize the training speed and classification accuracy of the model on the basis of the Resnet-18 network model. Firstly, the center loss function is introduced to optimize the softmax loss function, thereby improving the classification accuracy. Secondly, the original SGD optimizer is replaced with the Adam optimizer to address the slow training speed of the Resnet-18 network model. A garbage dataset is utilized to validate the effectiveness and feasibility of the proposed algorithm. The results demonstrate that after 40 rounds of training, the classification accuracy of the model optimized by multiple strategies increased from 81.42% to 87.42%. The neural network model optimized by multiple strategies can effectively identify classification problems, thereby enhancing the accuracy and stability of the model.