The Convergence Batch Dataset Algorithm Based on Deep Learning Model

Abstract

In the training process of deep learning model, GPU is basically used for accelerated training. Batch is a key concept in accelerated training in deep learning. By adjusting and combining the samples in the batch, we seek to ensure that the sample combinations of batch are different and that the sample labels of each batch are the same at each training iteration of the depth model. The algorithm is called CBDA (Convergence Batch Dataset Algorithm). Although the algorithm sacrifices a certain amount of computing time and enlarges the number of iterations, it delays the increase of fitting and improves the generalization of the model. Based on the accepted MNIST dataset, the experimental results confirm the advantages of CBDA.