Research On Domain Adaptation And Task Porting Methods For Transfer Learning

Transfer Learning refers to learn new knowledge utilizing learned knowledge based on their correlation.According to whether applying deep learning methods,transfer learning can be divided into two types: traditional transfer learning and deep transfer learning.It can be called domain adaptation when the feature space and label space of the source domain and the target domain are the same,but marginal distribution and conditional distribution are different.Transfer learning aims to solve the problem of how the machine learning task utilizes outer data to learn efficiently when the samples and the labels are lack.Therefore,transfer learning is of significant value to real applications.I finish the following works all by myself.As to domain adaptation in traditional transfer learning,a domain adaptation algorithm,Manifold Joint Distribution based on Optimal Transport,is proposed,which aims to deal with distorted features and divergent distributions between the source domain and the target domain simultaneously.This algorithm converts samples to Grassmann space by manifold feature transformation and optimizes optimal transport coupling matrix and classification function concurrently.The experiments on digits recognition and face recognition datasets demonstrate that comparing with baseline methods,this method is significantly efficient.As to the negative impact of outlying samples to transfer learning,a strengthening algorithm of domain adaptation based on remapping the outlying samples is proposed.It means that those outlying samples should be mapped first,and then they are used for transfer learning together with those non-outlying samples.The experiments launch on the text dataset of sentiment classification and image datasets.And the results state that this algorithm is universal,efficient,and robust.As to deep transfer learning,from the perspective of reusing the entire network,research on taking the entire semantic segmentation network as the preprocessor of the classification network to explore whether the segmentation network benefits the classification network is launched.The results show that this path is helpless to the target task in most scenarios.The domain adaptation algorithm proposed in the paper can be applied in real applications,such as image digits recognition and face recognition.The strengthening algorithm of domain adaptation can be regarded as a preprocessing method of samples to improve the performance of extant domain adaptation methods.The research about task porting in this paper can be the foundation of future task porting methods.