In the process of deep learning network training, most existing methods aim to improve the model effect focus on the network. However, to improve the effect and accuracy of the model it is necessary to pay attention to the characteristics of the data. In this paper, batch-attention, a new training framework for deep learning model, is proposed, which changes the original training method from the data level. It is shown that the method can coordinate overfitting and underfitting of the deep learning model. Experimental comparisons using Resnet34, TNT and efficientnet-b7 on Cifar10 and Cifar100 data sets respectively prove that the batch-attention model has improved both accuracy and F1-score in the test set compared with the benchmark model. In addition, the mechanism of batch-attention is further analyzed in the follow-up experiment.

deep learning  /  overfitting  /  attention mechanism  /  supervised learning  /  machine learning