A Multi-Component Heterogeneous Transfer Learning Approach For Image Classification

Traditional Machine Learning(ML)models work well under an assumption: the distribution of training and test data are the same.However,this does not hold in many real-world applications.Fortunately,Transfer Learning(TL)has alleviated this problem effectively,which relaxes the limitations in traditional ML models and transfers knowledge from Source Domain(SD)to help in learning models in Target Domain(TD).In addition,the proposal of Heterogeneous Transfer Learning(HTL)widens the application scope of traditional ML further,benefiting from its reusing knowledge among heterogeneous domains.In HTL,the SD components which are distributed most similarly to TD,are expected to function as a conduit connecting SD and TD to facilitate transferring knowledge.However,most of HTL methods haven't paid enough attention on the potential effect of SD components in knowledge reusage.If SD is dominated by the irrelevant or counterproductive components,the performance of TL will be severely decreased.Aiming at this issue,we firstly made analysis on the SD components.Based on this,we proposed the Multi-Component Heterogeneous Transfer Learning(MCHTL)algorithm and applied it in image classifications.The main works are summarized as follows:(1)We introduced the concept of component in HTL and made analysis on the effects imposed by SD components on knowledge transferring.And the analysis results showed: the effects imposed by different components in SD in TL were discriminative and the existence of some components would degrade the performance of TD models;(2)Based on this finding,we proposed a new approach that tried to reduce the impact of the counterproductive components by weighting the components.Specifically,we firstly split the SD into discriminative components,on which we trained corresponding models for TD.Further,we utilized weights to measure the performance of each model and adjusted these weights to optimize the final model.The experimental results on four benchmark datasets showed the superiority of MCHTL over the-state-of-art related methods.