Unsupervised Domain Adaption Learning Based On Extreme Learning Machine

As a neural network model with fast training speed and small number of manual parameters,the extreme learning machine has obtained extensive research and attention.Like all machine learning classification models,traditional extreme learning machines cannot classify accurately when test samples and training samples distributed differently.And the existing domain adaptation extreme learning machine models are still difficult to achieve good cross-domain classification performance.To this end,this paper builds extreme learning machine classification models with better domain adaptability from the aspect of classifier adaptive and feature representation adaptive.The main research contents are as follows:(1)The domain transfer based on the ELM parameters often cannot work well when the target domain is unsupervised.And the existing ELM based cross-domain classifiers only adapt the ELM output weights but ignore the adaptive learning of the ELM hidden layer weights.For the above problems,an unsupervised domain adaption classifier via extreme learning machine is proposed.First,the minimum error reconstruction learning is used for all source domain data and target domain data through the ELM based autoencoder(ELM-AE).The ELM-AE parameters can simultaneously represent source and target domain data through reconstruction learning.Then,we set the ELM hidden layer weights to the transposition of the ELM-AE reconstruction learning parameters,which can make the ELM hidden layer weights have better domain invariant characteristic.Finally,we embed joint probability distribution matching regular term and manifold regular term in the objective function of solving the output of the extreme learning machine.And we can finish the domain adaption adjustment of the output weights without the label information of target domain.(2)The parameters of the domain adaptation extreme learning machine need to have both classification function and domain adaptation function,which often cannot be simultaneously fully utilized.In addition,traditional matching method based on data probability distribution cannot find common subspace of the source and target domains under large difference between domains In order to alleviate the pressure of double functions of classifier parameters,the entire ELM learning process is mainly divided into two stages: feature representation and classifier adaptive learning,thus a joint feature representation and classifier learning based unsupervised domain adaption ELM model is proposed.First,in the stage of feature representation,we map source and target domain data to their respective subspace while minimizing the difference in probability distribution between the two domains.Next,the geometric differences between source and target domain subspaces can be further reduced by using the idea of subspace alignment.Then,in the stage of classifier adaptive learning,the target domain smooth manifold regularization term is used to improve the parameters adaptive ability.Finally,when optimizing the objective function,we can use the idea of EM algorithm to gradually optimize the parameters of the two stage until convergence.The experimental results on datasets of USPS vs Mnist,Office vs Caltech-256 and MSRC vs VOC2007 show that the proposed models can effectively solve cross-domain classification problems and have better domain adaptability.