The Study In Support Vector Machines And Deep Learning Algorithm

In this thesis, the Support Vector Machines method, the implementation of face recognition system and the deep learning method are deeply studied. while the Support vector machine and deep learning are both novel data classification method based on the statistical learning theory. And they have attracted great attention in the field of machine learning and pattern recognition. The main structure of this thesis is as following:Firstly, we proposed a novel robust modified support vector machines method. For the reason that the traditional SVMs method are very sensitive to the outliers, especially when constructing the support vectors and in some times it will lead to the decreasing of the performance, we proposed a new modified SVMs method to solve this problem. We first take use of the majority of the data distribution. Then we apply the distance between the two classes and modify the original SVM objective function. In the last step we search the best weights to make a balance between the distance and the maximum margin. We developed an improved method for optimizing over the SVM algorithm. Experiments on data sets clearly demonstrate improved performance and this method is also readily applicable for data sets with different distribution.Secondly, a face recognition system is implemented based on the methods of Gabor Filter, Principal Component Analysis and Support Vector Machines. We use the Gabor Filter to preprocess the photos. And then take use of the PCA to get the major features; in the end we use the SVMs to classify the dates. And we use the ORL to test our system. The result shows that our method can get good performance.Thirdly, combined with Support Vector Machines and Stacked Sparse Auto Encoder a new classifier method is proposed in this part. The Stacked Sparse Auto Encoder is a kind of deep learning method. It can learn the features from the datasets. This method has better performance than the traditional method. And then the experiment on the MNIST shows that our method is better than the other classical method such as SVM, BP, KNN and so on.