Multiply Connected Covering Learning Algorithm And Its Application

Covering learning is an important constructive learning method in the machine learning field, it has attracted much attention for a long time. In this paper, we used the connectivity of Lie group space as the foundation to solve the core scientific problems of characteristics optimization during covering learning. The main work in the last three years is as follow:1) We mapped the research objects with different category characteristics into the space of multiply connected Lie group, so as to present the category information of images by employing its multiple-valued representation, and then proposed multiply connected covering learning algorithm.2) In order to solve the path selection problem in multiply connected covering learning algorithm, we gave the method to calculate the weights of roads’ cut points and selected the road which consisted of cut points with the highest weights as the relatively optimal path expression to optimize the algorithm. Thus, we proposed weighed multiply connected covering learning algorithm.3) In order to solve the path intersection problem in multiply connected covering learning algorithm, we introduced the idea of Fisher projection to optimize the algorithm, and calculated the optimal projection direction for projecting roads on the connected space to increase road density of each connected space. Thus, we proposed multiply connected covering learning algorithm based on Fisher projection.4) We applied the above algorithms to pattern classification of the image library and handwriting recognition, the effectiveness of the algorithms was verified by experiments.In summary, on the one hand, the research of this article formed multiply connected covering learning algorithm innovatively, which provided new solutions for problems of characteristics optimization during covering learning. On the other hand, the validity of the algorithms proposed in this article was verified by experiments, which provided the application background for these methods.