Research On Cross-domain Lifelong Machine Learning Algorithm

With the advent of the era of cloud computing,data grows explosively and the problems of vast amounts of data size,variety of data types,and low density of data value stand out increasingly.Otherwise,in a lot of practical machine learning situation,the problems such as lack of training samples and high cost of artificial tagging still remain.These problems bring big challenges for traditional machine learning.For example,the running time becomes longer and the prediction performance becomes worse.Therefore,humans expect that the machine learning system is able to store knowledge like our brain and when new task arrives,the system are able to use existing knowledge to improve the learning of new task and update existing knowledge after learning.Under such background,relevant researchers imitate the human brain learning mechanisms and propose lifelong machine leaning.Lifelong machine learning system is the learning system that is capable of efficiently and continually dealing with tasks,transferring knowledge and refining knowledge.In real life,we human beings are faced with various cross domain tasks such as cross scene object recognition,and our brain is capable of transferring knowledge,refining knowledge,updating knowledge and then well accomplishing such cross domain tasks.So,dealing with cross domain tasks is essential for an ideal lifelong machine learning system.However,existing lifelong machine learning algorithms don't focus on cross domain tasks and when faced with cross domain tasks,the existing lifelong machine learning system is not robust enough.To this problem,this paper proposes a lifelong machine learning algorithm based on manifold embedding and a lifelong machine learning algorithm based on auto encoder as follows.(1)This paper proposes a lifelong machine learning algorithm based on manifold embedding.The proposed algorithm generates the geodesic flow by embedding Gaussian manifold to task subspace,chooses optimal subspace from geodesic flow and performs task projection before model learning.The proposed algorithm is robust for cross domain tasks and runs efficiently.(2)This paper proposes a lifelong machine learning algorithm based on auto encoder.The proposed algorithm uses the encoder weight parameter of the past task to improve the optimization of the new task.To be specific,we constrain the encoder weight parameter of the current task and the one of the past task are similar to a certain degree.The proposed algorithm is able to extract feature representation which is robust for domain shift.In order to verity the robustness of our proposed algorithm for cross domain task,we construct three cross domain multi-task datasets.Then we conduct enough experiments and the experiment results show that our proposed lifelong machine learning algorithm based on manifold embedding and lifelong machine learning algorithm based on auto encoder achieve better recognition accuracy on cross domain multi-task datasets than the existing lifelong machine learning algorithm.On prediction accuracy,the proposed lifelong machine learning algorithm based on auto encoder performs better;On running efficiency,the proposed lifelong machine learning algorithm based on manifold embedding performs better.