Investigating The Problem Of Robustness Recognition Based On Deep Transfer Learning

With the prosperous of machine learning,deep learning methods,as an important branch of machine learning,have made impressive achievements in many Artificial Intelligence fields,such as Automatic Drive,Speech Recognition,Image Detection and Recognition,Image Style Transfer,etc.To achieve desirable results,training complex deep neural networks with high performance in these fields usually requires a large amount of labeled training data.However,for some new domains and tasks,collecting and annotating datasets are extremely expensive and time-consuming processes,and for some time-related domains and tasks training data is easy to be out of date.Training neural network model in these domains and tasks is difficult due to the lack of data.Therefore,transfer learning,as an effective method to solve such problem,has attracted the attention of researchers and developed fast recently.Transfer learning is a new machine learning method that uses existing knowledge to solve problem in different but related fields.This research aims to study the transfer learning methods based on deep convolutional neural networks,and proposes new robust image recognition methods for both small-data transfer learning task and unsupervised domain adaption task.Our contributions are as follows:(1)To tackle the problem of lacking of flexibility in the construction of target network model in existing convolutional feature transfer methods,we propose a new convolutional feature transfer architecture based on the current high-performance deep convolutional neural networks,and three common used convolutional networks were chosen as examples to present the specific feature transfer process.Based on the above method,an adaptive feature input method is also proposed so that the convolutional layer of target network can focus only on the features that benefit the current layer's feature extraction.In addition,as a special case of the above method,we also propose a method that transfer feature and finetune parameters simultaneously for the case that the target convolution network and the source feature generator share the same network structure.(2)In the unsupervised domain adaption task,we innovatively introduce the maximum mean discrepancy penalty into the newly proposed Maximum Classifier Discrepancy method to match the inter-domain classification boundary and inter-domain feature distribution simultaneously,which facilitate the improvement of unsupervised domain adaptation.Besides,based on the characteristic of the Rectified Linear Unit,we propose a new regularization method for the rectifier convolutional neural network which is used in our method and is the most popular convolutional neural network recently.And,our regularization method claims more robustness accordding to the experimental results which shows that our method is more effective than the traditional regularization method in unsupervised domain adaption task.