Improving The Training Of Convolutional Neural Network Using Inter-Class Distance And Between-Class Distance

As an important branch of deep learning research,Convolutional Neural Networks have performed very well in image recognition and classification tasks in recent years.It has also been widely used in many fields to make people's lives more intelligent.Compared with exploring how to apply Convolutional Neural Networks in a wider range,understanding Convolutional Neural Networks is more important.Only by understanding Convolutional Neural Networks can we know the network structure clearly,then improved training methods that could improve and optimize Convolutional Neural Networks will be proposed.The research on multi-classification tasks in this paper includes two points.The first one is to evaluate the training process of Convolutional Neural Networks from the perspective of distance measurement.The second point is to improve the training of Convolutional Neural Networks based on the analysis of distance measurement to make full use of training data and make classification more accurate.Three indicators including C-index,DB-index,and between-class distance are used to evaluate the Convolutional Neural Networks in this paper.Combining the distance relationship between the features,we summarized the change characteristics of the distance measurement in different layers as the number of training epoch changes during the training process.Then the change characteristics will be used as a basis to improve the training of the Convolutional Neural Networks.To improve the training of Convolutional Neural Networks,the general method is to adjust parameters randomly.But adjusting parameters requires enough experience,and the combination of parameters is full of randomness,which can't guarantee the full use of training data.To avoid the time consumption caused by finding parameters randomly,in this paper we proposed a method to make the classification more accurate by analyzing the distance measurement of the deep features of the training data.In the classification task,the distance measurement between classes is closely related to the classification accuracy.Low classification accuracy means that the training of certain classes is not sufficient.The improved method proposed in this paper divides the training into three steps,including basic training,special training,and global training.Using distance measurement from basic training will help us detect underfitting classes,which stand for classes that are not trained sufficiently.Special training is performed on the underfitting classes based on the initial model obtained from basic training,so that the model can strengthen the learning of underfitting classes.Special training models are produced in this step.Data from all the classes will be trained based on special training models in the global training process.Finally,a model with higher classification accuracy will be produced.This method uses training data in multi-class classification in a more efficiently way,which also improves classification accuracy.Experiments show that the improved training method proposed in this paper is not only suitable for simple data sets,but also performs well in the training of relatively complex data sets and networks.