Kernelized Unified Domain Adaptation on Geometrical Manifolds

Abstract

Primitive machine learning algorithms like the k-nearest Neighbor (k-NN) and Support Vector Machine (SVM) are a major challenge for expert and intelligent systems that recognize objects with large-scale variations in lighting conditions, backgrounds, color, size, etc. The variations may be due to the fact that the training and test data may come from related but different domains. Considerable effort has been put into the advancement of domain adaptation methods. However, most of the existing work only concentrates on considering only a few of the following goals or objectives: (i) subspace alignment; (ii) Minimization of distribution divergence by using the Maximum Mean Discrepancy (MMD) criterion; (iii) Preservation of source domain discrimination information; (iv) Preservation of original similarity of the data samples; (v) Maximization of target domain variance. Current approaches to preserve source domain discriminant information can easily mis-classify target domain samples which are distributed near the edge of the cluster. In order to overcome the limitations of existing domain adaptation methods, and expert and intelligent systems, we propose the Unified Domain Adaptation on Geometrical Manifolds (UDAGM) framework. UDAGM optimizes all the aforementioned objectives jointly as well as uses the Regularized Coplanar Discriminant Analysis (RCDA) method for better inter-class separability and intra-class compactness. In addition, we extend our proposed framework UDAGM to a kernelised version in order to deal with non-linear separable datasets. Extensive experimentation on two real-world problems datasets (PIE face recognition and Office-Caltech) has proven that the proposed frameworks outperform several approaches to domain adaptation.