A Method Of Repairing Convolutional Neural Network Model Feature Loss Based On Feature Differencial Analysis

Artificial intelligence,especially deep learning,has become an indispensable part of modern computer systems.In recent years,Convolutional Neural Networks(CNN)have achieved great results in image classification and have been widely used in various fields of computer vision.Like traditional software,Convolutional Neural Networks also have bugs,result-ing in poor classification(prediction)accuracy of convolutional neural networks.How-ever,it is impossible to repair the bugs of the neural network by directly modifying the model like fixing the bugs of traditional software.The existing methods to solve the bugs of neural networks are basically through additional training data.However,due to the poor interpretability of neural networks and the difficulty of guaranteeing the quality of additional training data,the methods of fixing bugs have limited effectiveness.The feature loss is a common bug in image classification of Convolutional Neural Networks.Inspired by Res Net,we propose a method to analyze the feature difference between images with feature loss and its variant images in each layer of the neural network model,and then add Shortcut Connection to convolutional neural networks to solve the bug of feature loss in Convolutional Nerual Networks.Through 18 experiments on the three datasets(CIFAR-100,SVHN,GTSRB)and the two models(Res Net18 No Shortcut-Connection,VGG-11),it is proved that the meth-od we proposed is effective in fix the feature loss bug.Take the debugging results of the CIFAR-100 data set on the VGG-11 model as an example.After adding a Shortcut Connection,the top-1 test accuracy of the original model increased from 65.01% to67.24%,and the feature loss bug fix rate was 62.7%.Add after two Shortcut Connec-tions,the top-1 test accuracy of the original model increased from 65.01% to 69.54%,and the feature loss bug fix rate was 79.3%.