Kernelized Unified Domain Adaptation on Geometrical Manifolds

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