A Study On Multi-Label Transfer Learning Algorithm And Application In The Bird Sounds Recognition

Transfer learning is a kind of machine learning algorithm,which is proposed to solve the problem of cross domain learning.Transfer learning algorithms transfer knowledge through common latent feature or structure between different but related domains.Early transfer learning algorithms assume that each instance is only related to one label.But each instance can be simultaneously labeled with multiple categories in image,video and text classification problems,and traditional transfer learning algorithm is no longer applicable with it.Therefore,some studies focus on multi-label transfer learning in recent years,which solve the multi-label transfer learning problem by reconstruct multi-label data into single-label data or adapt existing single label transfer learning algorithm.Current multi-label transfer learning research is all based on feature transfer method that transfer feature distribution or label correlations from source domain to target domain by constructing a common latent feature representation.However,few studies on the effect of the difference in the feature distribution or label correlations between domains.Therefore,this paper study on the effect of the difference in the feature distribution and label correlations between domains in multi-label transfer learning problem.The main contributions of this paper include:(1)According to the problem of the different feature distribution between domains,propose a new multi-label algorithm for domain adaptation named Multi-Label Transfer Learning via Maximum mean discrepancy(M-MLTL).Existing approaches ignore the different probability distribution between the source and target domains,then source domian data can not exert its effect.To tackle this problem,we propose a novel algorithm for domain adaptation named M-MLTL.M-MLTL is a transfer learning algorithm based on feature transfer,which map source and target domain samples into a shared latent feature subspace.Then M-MLTL introduce Maximum Mean Discrepancy to ensure that the samples are similar in the subspace.Thus,effect of the difference in the feature distribution is avoided.Experimental results on two image classification data sets show that the new algorithm has higher classification accuracy and computational efficiency compared with the existing similar algorithms.(2)According to the problem of the different label correlations between domains,the theory of local label correlation in multi label transfer learning is proposed.Then we propse Multi-label Transfer learning using LOcal Correlation(MTLOC).Existing multi-label transfer learning algorithms employ the global label correlations.However,the global label correlations from source domain cannot be directly applied in the target domain due to the discrepancy between the source domain and target domain.We propose a novel multi-label transfer learning algorithm named MTLOC,which extend a multi-label learning algorithm named "Multi-Label learning using LOcal Correlation(ML-LOC)" into transfer learning algorithm.MTLOC obtain the shared label correlations by minimizing the weighted loss function.Then use shared label correlations as auxiliary features in the target domain classification.Experimental results on two image classification data sets show that the transfer learning algorithm based on local label correlations has higher classification accuracy compared with the transfer learning algorithm based on global label correlations.(3)Propse a new Recognition method of Multiple Bird Species in audio recordings based on Feature Transfer(FT-RMBS)that provide a solution for recognition when lack training samplesTo deal with the problem of inadequate sample in multiple bird species recognition,propse a new recognition method of multiple bird species in audio recordings based on feature transfer(FT-RMBS).Which try to use bird sounds of single species to train a multiple bird species recognition model.FT-RMBS map original audio feature of birdsong into latent audio feature in that single and multiple species sample have similar distribution.The new recognition method provide a solution for recognition when lack training samples.Experimental results on multiple species bird sound data sets show that the new method has achieved good recognition results on 4 global evaluation metrics.Experiment on artificial bird sound data sets shows that FT-RMBS has 20%higher recognition rate than other recognition method,which without feature transfer.