GFK And Its Application In Domain Adaptation

Domain adaptation is a new filed in machine learning which is associated with transfer learning.Unlike traditional machine learning methods,it is assumed that the training set and testing set are not identically and independently distributed.On this basis,the training set is named source domain,and testing set is named target domain.Source domain and target domain are similar to each other.Domain adaptation is mainly applied to text analysis,machine translation,and statistical pattern classification.In recent years,it was applied in the field of computer vision and has become a hot research topic.This paper studies the application of GFK in domain adaptation.The main contribution of the paper is as follows:1.An new algorithm for domain adaptation based on SGF is presented.First,the generative subspaces of source domain and target domain are extracted,and they are treated as two points on the Grassmannian manifold.Then,geodesic flow curve connecting the source and target domains on the Grassmannian is constructed,a fixed number of points on this curve are sampled and used as projection matrix for the original feature space.The original feature vector is projected into subspaces using these projection matrixes and these projection vectors are concatenated into a new feature vector.Then,the dimensionality of the new feature vector is reduced by PLS method.Finally,the resulting feature vectors are used to construct classifiers.In this paper,we use the cosine classifier.Finally,we apply our algorithm to object recognition task and face recognition task.One database for object recognition and two databases for face recognition are used to test the algorithm.The experimental results show that our new algorithm outperforms the conventional SGF algorithm in object recognition.Moreover,using this algorithm,we get better performance in face recognition.2.Kernel PCA algorithm for domain adaptation based on GFK is presented.First,geodesic flow kernel(GFK)is constructed using the source and target domain.Then,for given feature vectors,new feature vectors are extracted by KPCA algorithm in which the GFK is used as the kernel function.Finally,cosine classifier is used for recognition task.Again,we apply the algorithm to object recognition and face recognition task.Three databases are used to verify our new algorithm.The experimental results show that this algorithm outperforms unsupervised GFK method in objection task,and moreover,weapply it in face recognition task and get better performance.