Research On Critical Technologies And Applications Of Extreme Learning Machine-based Incremental Learning

Incremental learning algorithm is an effective method to achieve how to update the model with the growth of data in an online manner.In the meantime,the emergence and developments of Extreme Learning Machine(ELM)provide a new idea for rapid mod-eling and updating.Incremental learning and ELM have made great achievements in the past few decades,but there is still certain distance from practical application under the background of the rapid developments of data.There are three problems in the existing methods.Firstly,ELM-based incremental learning needs to determine the network struc-ture beforehand while dealing with data increment.However,the continuous increase of data will lead to a problem of deciding appropriate network structure.Secondly,ELM-based incremental learning algorithms aim at the incremental updating brought about by the changes of data,but data might exhibit the feature of imbalance during the increment along time,or even concept drift in the process.Thirdly,the existing incremental learning researches based on extreme learning machine focuses on the model updating caused by simple data increase or decrease,but the change of data is more complex in the specific application scenario,which usually presents a mixed situation of multiple data increments,thus making the corresponding model updating more complex.Aiming at these problems,the thesis focuses on ELM-based incremental learning al-gorithms,especially online learning algorithms.On the basis of summarizing the related algorithms of incremental learning and ELM,the thesis puts forward corresponding solu-tions to the limitations of current methods,and achieves the following research results:(1)An online sequential incremental extreme learning machine is proposed to syn-chronously perform structural increment during the process of data increment.Different from simple ELM algorithms based on data increments,the proposed algorithm monitors the current classification/regression error changes every time after model is updated with the newly generated data.When the errors change too much,the algorithm increases net-work nodes in the process,and uses the rank-one updates of the generalized inverse of block matrices to optimize the solution.Experimental results on classification/regression datasets show that the proposed algorithm has better classification/regression performance than other ELM-based incremental learning algorithms.(2)A weighted domain transfer ELM and its corresponding online learning algo-rithm are proposed to deal with the problem of imbalanced data learning,and applied to the problem of gas sensor drift compensation.To tackle with the concept drift and data imbalance generated in the increments of data,the thesis introduces weighted learning in domain adaptation algorithm to construct classification model.On this basis,the the-sis targets on the unlabeled data increment and derives the corresponding online learning algorithm.Experimental results on gas sensor data show that the weighted domain trans-fer ELM can achieve higher classification accuracy with less labeled samples,and in the meantime,its online learning version enables the model to perform online updates while maintaining the property.(3)Two online domain adaptation ELM models are proposed to deal with the online updates of semi-supervised learning model in the dynamic changes of data,and applied to gas sensor drift compensation.In the framework of semi-supervised learning,the changes oflabeled and unlabeled data include their own increments or decrements and the transfor-mation from the unlabeled to the labeled.Aiming at the online updates of the model result-ing from the series of changes,the thesis bases on two different semi-supervised learning model assumptions and proposes source domain-based and target domain-based online do-main adaptation ELM algorithms,respectively.Experimental results on gas sensor data show that both algorithms can perform model updating in an online learning manner,in which the source domain-based online learning model has higher classification accuracy when labeled samples are fewer,while the target domain-based online learning model gains better classification ability with the increase of labeled samples.