A New Neighbor Reconstruction Method And Its Application In Machine Learning

In recent years,artificial intelligence has made a great breakthrough,and caused great attention in academic circles and industry,even attracted the attention of national policy.Artificial intelligence techniques have various applica,tions including computer vision,language and image processing,medical diagnosis,robot control,face recogni-tion,finance and economics,healthcare,and education,etc.In particular,one of the most well-known results of artificial intelligence is AlphaGo developed by Alphabet Inc.'s Google DeepMind,which has ever beat a human professional Go player.The crucial content of artificial intelligence is machine learning algorithms.Ma-chine learning is a multidisciplinary subject,which studies how computers simulate or achieve human learning behaviors in order to acquire new knowledge or skills and to reorganize existing knowledge structures to continuously improve their performance.Subspace method is a kind of important algorithm in machine learning.The subspace method uses the neighbor subspace to describe the corresponding sample points and the neighbor subspace is also used as the intrinsic parameters of the corresponding samples,and finally,the intrinsic parameters are used for dimensionality reduction and recon-struction of the samples.The classical subspace methods include local linear embedding(LLE),Laplace embedding(LE)and so on.At present,most subspace methods mainly focus on how to use subspace to describe the corresponding samples and extract the intrinsic parameters.For example,the LLE method utilizes the least-squares represen-tation coefficients of the neighbor samples corresponding to the center sample as the intrinsic parameters of the center sample.In this thesis,we present a novel and compact neighbor reconstruction method(NRM).This is a unified pre-processing algorithm for graph-based subspace methods.The proposed algorithm is implemented by iteratively linearly interpolating the center point and its corresponding neighbor points.The NRM can generate more neighbor points.Compared with the original neighbor points,the newly generated neighbor points can better describe the subspace structure of the center point and then more effective intrinsic parameters can be proposed.Based on the NRM,many subspace-based nonlinear feature extraction and feature selection algorithms can be significantly improved.In particular,the NRM is embedded in several classical algorithms such as LLE,Laplace Eigenmap(LE)and Unsupervised Feature Selection for Multi-cluster Data(MCFS)and has achieved 7%,2.6%,2.4%classification improvement on ORL,CIFAR10 and MNIST datasets accordingly.We also apply the NRM to the single image super-resolution algorithm(e.g.A+),and has achieved remarkable performance gains of 0.12 dB.