| With the fast development of computer technology and the widely use of digital products,the Internet is now filled with a huge number of images.In many scenarios,we need to do classification for these images.However,it is impossible to complete this task manually,which is the main reason for us to seek help from high-performance computers.On the other hand,deep learning has been one of the most significant breakthrough in recent years,and auto-encoder,as one common models in deep learning,has been proved to be effective in dealing with the problem mentioned above.Auto-encoder is a probability-free method,which could extract effective features from original.It has aroused widespread interest and has been applied in a variety of areas including object recognition,image classification,face recognition and nature language processing.The main contribution of this paper by using auto-encoder for image recognition are as follows:1.A Hessian regularized sparse auto-encoders method is proposed.This method employs Hessian regularization to well preserve local geometry for data points.It can also efficiently extract the hidden structure in the data by using sparsity constraints.To evaluate the effectiveness,we construct extensive experiments on the popular datasets including MNIST and CIFAR-10 dataset and compare the proposed HSAE with the basic auto-encoders,sparse auto-encoders,Laplacian auto-encoders and Hessian auto-encoders.The experimental results demonstrate that HSAE outperforms the related baseline algorithms.2.A large-margin auto-encoder method is presented.The large-margin constraints could increase the discernibility degree of hidden layers.Besides,it maintain a minimum distance among examples with different labels under the circumstance of reducing distances of congener samples.To put it in practice,we only consider the nearest neighbors of each instance and therefore creating a safe boundary from data of different classes.As a result,we build a classifier with better performance. |
PDF Full Text Request