Research On Classification Method Of Multi-source Domain Knowledge Transfer Learning

Classification technology has a very wide range of applications in computer vision,medical diagnosis,natural language processing,dialogue systems,behavior recognition and other fields,and it is one of the important research directions in the field of artificial intelligence.As an important branch of machine learning,multi-source domain knowledge transfer learning has received great attention from researchers in recent years.The classification method based on multi-source domain knowledge transfer learning can effectively imitate human "analog" behavior,and classify and judge things in new fields by acquiring knowledge in different fields.Classification and judgment through multi-source domain knowledge transfer learning does not require a large amount of data and labels,and there is no relationship between the data.This feature can not only improve the learning efficiency while ensuring data security,but also be more conducive to the application of the model.Therefore,multi-source domain knowledge transfer learning has become one of the research hotspots to solve the problem of classification technology.This paper conducts relevant research on common problems of multi-source domain knowledge transfer learning at this stage and proposes corresponding solutions.Aiming at the problem of under-fitting of the model caused by the large individual differences of samples in multiple domains in multi-source domain data samples,a multi-source domain transfer learning classification method based on support vector machine is proposed.This method re-weights the sample data of the source domain by using the edge probability difference,and then recalculates the maximum mean difference distance between each source domain and the target domain according to the weighted samples,so as to reduce the individual differences between the sample data of the source domain and the target domain;In order to reduce the computational complexity caused by the reweighting of each sample in the domain,approximate pole support vectors are introduced into the algorithm.The approximate pole support vector greatly reduces the size of the training samples and improves the model efficiency by finding the training samples near the super-large plane.Finally,the objective function for the final classification is deduced according to the structural risk minimization theory,and the objective function is confirmed by theoretical proof.The solution process is a quadratic programming problem with an optimal solution.Aiming at the problem of poor model adaptation caused by different feature distributions of samples in multiple domains in multi-source domain data samples,a multi-source deep transfer learning classification method based on feature alignment is proposed.The method uses the maximum mean difference distance to obtain the similarity weight of each source domain sample and the target domain sample,and then re-weights each source domain sample according to the similarity.Then use the semi-supervised manifold alignment method to map the common features of the source domain and the target domain,the unique features of the source domain,and the unique features of the target domain into a low-dimensional manifold space,which can ensure that the source domain data remains original.In some manifold structures,similar data are still similar in low-dimensional space,and irrelevant data are still irrelevant.By reducing the intermediate subspace,the amount of calculation is reduced and the model operation efficiency is improved.Then,the conditional probabilities of the source and target domains are constrained by convolutional neural networks and maximum mean difference to avoid negative transfer.Finally,an effective classifier ensemble selection strategy is designed and implemented,and the experimental results confirm that the method can effectively classify data.This method is used for multi-source domain data classification,which solves the problem of poor model adaptation caused by different distribution of data samples.Aiming at the problem that the relationship between multi-source domain data samples is not fully utilized,resulting in large differences in class balance,a multi-source domain deep transfer learning classification method based on class adaptation balance is proposed.The method considers the difference between the conditional probability and the marginal probability between the source domain and the target domain at the same time,and introduces a balance factor to adjust the proportion of the influence of the conditional probability and the marginal probability,so as to achieve the class balance between the source domain and the target domain.Before class balancing the source and target domains,first map the source and target domains to the same target space,and then align multiple samples from the source and target domains.The data in the source domain is dimensionally reduced.Then,a dynamic balance factor is introduced to adjust the conditional probability and marginal probability of the data samples to achieve a better classification effect.Finally,a regularization term is introduced to prevent the objective function from overfitting.Compared with a single classifier,the classifier ensemble obtained by the experimental training has better classification effect and higher stability.Aiming at the problem that multi-source transfer learning has high computational resources and memory overhead in actual mobile terminal applications,and mobile terminals cannot be used effectively,a multi-source mobile transfer learning classification method based on dynamic model compression is proposed.This method first pre-prunes and compresses the source domain model by pruning the BN channel,and then calculates the classification loss and the loss of the maximum mean difference on the mobile terminal,and calculates different source and target domain classifier bands on the server side.to the classifier loss.After that,the model parameters are updated on the server side,and the source domain model parameters are returned to the corresponding client.Under the premise of ensuring the use efficiency and accuracy,the method greatly reduces the calculation amount of the mobile terminal and improves the system response speed.At the same time,the server and client transmit the model after transfer learning instead of user data,which effectively improves the data security in practical applications.