Research On Lifelong Machine Learning Algorithms Based On Deep Learning Representations

Nowadays,with more powerful computing capabilities and the explosive growth of data today,the traditional machine learning has been unable to meet the needs of the development of new artificial intelligence while lifelong machine learning has been receiving more and more attention.Lifelong learning machine can be applied in many fields,such as computer vision,data mining,medical diagnosis,and automatic robot.The main work of this paper is to solve the problems of insufficient label data,high input dimension,and inconsistent distribution of tasks,which are difficult to solve in traditional lifelong machine learning systems.Combining with the deep learning representations,the following two innovative lifelong deep learning models are proposed:1.Hierarchical lifelong learning algorithm(HLLA)with a representation shared by all tasks.The feature representation can be any deep learning feature model that uses the back propagation algorithm to update its parameters.The parameters of the inference function are optimized based on an efficient lifelong learning algorithm(ELLA).Deep-shared networks can extract representative features from high-dimensional and unlabeled data,thereby greatly reducing the amount of labeled data required for learning new tasks.2.Grouping life-long deep learning model(G-HLLA model)which automatically learns a grouping for tasks and features are shared only in the same group.The feature representation is Deep Belief Network(DBN)constructed by Reconfigurable Boltzmann Machines(RBMs),and the inference function is based on the ELLA algorithm.The feature network obtained after grouping based on reconstruction error has better robustness and can better solve the problem of inconsistent task distribution.In order to verify the robustness of the proposed two methods for lifelong learning tasks,we also fully validate the two models by experiments.In the six commonly used lifelong learning databases such as landmine detection,animal classification,and student achievement prediction,comparisons are made based on several representative lifelong machine learning algorithms in recent years.The experimental results show that both models have significantly better performance than the baseline lifelong learning model in terms of classification accuracy and regression error.