Research Of Adversarial Learning In Neural Network Model Optimization

As a science springing up in the era of big data,machine learning methods are not only based on data but also confined to data.Good understanding of data is an indispensable part of an excellent model solution,furthermore,it can be a crucial means of solving issues in a specific scenario.In the case where data set have labels at multiple scales,neither the methods focusing on the single main task label nor unreasonable using of multi-scale labels will bring about detrimental performance because of attributes bias of data set.Unfortunately,this situation is common in machine learning matters.Basing on the analysis of multi-scale label issues,in this thesis we focus on the data set of pairing phenomena for a class of multi-scale labels,get the generalization of the problem of grouped or biased attribute of training data.As a result,we build a domain migration problem caused by a repulsive combination of attributes between training set and test set.After that,we progressively presented two solutions with neural network models as follows:Firstly,by using the characteristics of multi-scale attribute labels in samples,this thesis designs a property disentangle network model based on the adversarial learning mechanism based on the causal relationship between attributes,which uses the confusion between attributes on the main classification task branch.The branch reduces the interaction between the attributes,thus achieving the effect of improving the performance of the main classification task.Secondly,for the coupling relationship between attributes and the incomplete attribute decoupling,this thesis firstly utilizes a random combination of data augmentation methods for the single attribute vector to classify the training data set,and then based on the identity of the generated data,based on the confrontation Learning ideas construct different cost-backpropagation mechanisms for samples of different attribute combinations,thus achieving disentangling of different attribute components in the sample.In the experimental session,we prove that such a model can further promote the performance of the model in the main classification task by improving the disentangling level,and verify the effectiveness and practicality of the method on a real authentication data set.