Learning Of Radius-margin Based Model With Log-det Regularization And Its Application

As one of the most popular fields in machine learning, learning methods based on radius-margin bound have received extensive attention recently. The change of radius is hidden during feature transformation, which is is of great importance to the learning and optimization of classifier performance. However, existing radius-margin based methods have some deficiencies, thereby restricting the improvement of classification performance. Besides, current methods are not scalable to large-scale datasets due to higher computational complexity.To address the aforementioned issues, we proposed an efficient radius-margin based support vector machine with Log-det regularization term(L-SVM). Due to the non-convexity of traditional radius-margin SVM model, we firstly approximate the radius-margin ratio with relaxation, and then transform the original problem into an unconstrained convex one. To reduce the risk of over-fitting, Log-det regularization term is incorporated into the model and an efficient solution strategy is pr oposed. Experiemetal results on several UCI datasets for linear and kernelized classification validate the effectiveness of the proposed model. In order to further validate the generalization ability of our L-SVM model, we apply the proposed model to action recognition task. We firstly extract a variety of local features from videos, and encode them with Fisher Vector for mining the distribution of local features, then perform action recognition with L-SVM classifier. Results on benchmark datasets show that the proposed model has good generalization ability and can be widely applied to different classification and recognition tasks.By combining hierarchal feature extraction and classifier learning, convolutional neural network(CNN) has achieved excellent performance in many computer vision tasks. Inspired by this, we embed the proposed L-SVM model into CNN framework, and alternative optimization is proposed to improve the action recognition performance. Besides, in order to improve the discriminative power of CNN feature, we introduce a center based regularization term, minimizing intra-class divergence while maximuming inter-class divergence. Moreover, we also add more supervision information to the intermediate layers of CNN,which improves the convergence perfromance of the network. By comparing the performance of different radius-margin based learning models and investigating the effect of incorporating center based regularization term/ supervision information into the model, we show a new perspective of combining radius-margin based learning models with CNN.