Local Linearity Detection Based Research For Quasi-Linear Support Vector Machine

Object recognition and classification is one of the most fundamental and representative problems in the artificial intelligence field. The overcome of such problem will be significant and meaningful. By tons of researches and experiences, it is commonly believed that two factors are essential for the performance of object recognition and classification. They are abstract feature representation and algorithm to optimize a given problem. On the one hand, feature representation, as the only way for a computer to know the real world, determines the upper bound of the potential intelligent level. On the other hand, for a given feature representation and a specific situation, the effectiveness of an algorithm decides the approaching level of the potential upper bound. Therefore, feature representation and algorithm formulation are always two hot topics in the artificial intelligence field.In this paper, starting from the algorithm formulation, local linearity detection is implemented to improve a previously proposed model and gives it a thorough analysis and summarization. Based on it, a local linearity detection based quasi-linear support vector machine is proposed. Moreover, by combining the recent improvement of deep learning, analysis and simulation experiments are conducted for exploration of the suitable applications of the proposed method. The major outcomes and conclusions of this paper can be summarized as follow:Firstly, with a visualization of the gated basis functions of a quasi-linear kernel, the inside working principle is displayed for better understanding to readers. Then, several experimental results are presented to show the disadvantages of the previously proposed model, so as to express the meaning and value of the improvement method proposed in this paper.Secondly, a quasi-linear support vector machine(SVM), as one type of nonlinear kernel SVM, include a kernel formulation process. Different from commonly used kernel functions, a quasi-linear kernel has a unique formulation for a specific problem. Based on the experimental results, it can be concluded that the proposed kernel formulation method can effectively improve the performance of recognition and classification. Besides, a quasi-linear kernel is formulated by the detected local linearity, so it is robust to the influence of hyper-parameters. Furthermore, it can avoid the time-consuming process of cross-validation, which makes it more suitable for the real-world applications.Finally, deep neural networks based feature representation has been empirically proven to be effective to improve the artificial performance of a computer. Although such technologies reach state-of-the-art performance in various application fields, one factor behind it can not be ignored. That is a huge number of labeled data. However, in most of the real-world applications, it is hard to obtain a large labeled data set. Therefore, the training of a deep neural network is prone to overfit the data set. By combining the proposed method with auto-encoder and transfer learning, a better performance is obtained in this paper. By careful analysis, it can be concluded that the proposed method is suitable for the problems that only include a small set of labeled data.