Research On Training Method Of Convolutional Neural Networks Based On The Ensemble Of Receptive Fields

Translation transformation is a data augmentation method widely used in the image classification tasks.We analyzed the physical meaning of translation by mathematical method,and discovered that the translation can make Convolutional Neural Network pay attention to the features of the central area of the image during the training process,thus making the receptive field of convolutional neural network nonuniform.If the thickness is used to indicate the degree of attention paid to each region of the receptive field,the translation method makes the receptive field of convolutional neural network present a"pyramid"distribution with high middle and low margin in the training process.The resolution of each region of the receptive field of the human retina is different:the macula in the center provides the highest resolution of the entire retina,while the resolution of other regions is significantly lower than that of the central macula region.This inspired us to analyze the significant features contained in each region of the image in a general sense,and proposed the following assumption:compared with the marginal area,the central area of the image is more likely to contain significant features.We have verified this hypothesis through relatively complete experiments.Based on this assumption,we designed a new receptive field ensemble training method for convolutional neural network image classifier by using translation method.This training method divides the whole training process into several stages,each stage corresponding to different receptive fields and learning rates.Specifically,in the first stage of training,convolutional neural network focused on the features of the most central region of the image,while the learning rate was the highest.With the process of training,the areas payed attention by the convolutional neural network spread from the center to other areas,and the learning rate gradually decreased.We use this method to make convolutional neural network not only pay attention to the features of the central area of the image,but also take care of the features of the marginal area in the training process.We call this method the ensemble of receptive fields since it integrates multiple receptive fields in the training process.We used two different types of deep convolutional neural network and take experiments on two different image classification benchmark data sets.The results show that,on the premise of the same number of iterations during training,the training method of receptive field ensemble produces higher test accuracy than the that of single receptive field.Different from the improvement of general depth-learning image classifiers in recent years,our method is interpretable from principle to specific steps.Our work not only provides a new training method for convolutional neural network,but also may inspire the study of interpretable neural networks.