Research On Auto-encoder Based Representation Learning

As an important technique for realizing artificial intelligence,machine learning has attracted more and more attention from the academic and industrial circles.Features have an important influence on the performance of the machine learning model,and the representation learning is an important means to acquire features.This paper starts with the auto-encoder based representation learning algorithm,analyzes the existing auto-encoder representation learning methods,compares the advantages and disadvantages of the algorithms,and finally proposes our own auto-encoder based representation learning method.The main content of this paper is as follows:Firstly,the related knowledge of neural networks and auto-encoder are reviewed.The network structure,activation function,back propagation algorithm and training optimization algorithm of the neural network are mainly introduced.The network structure of the auto-encoder and several typical regularized auto-encoders representation learning methods are introduced in detail,and the characteristics of these regularized auto-encoders are analyzed.Secondly,aiming at the problem of what is a good representation learning method,metric auto-encoder algorithm is proposed.The algorithm combines metric learning ideas with an auto-encoder so that the feature representation learned by the auto-encoder can maintain the metrics information of the sample.And from the perspective of Bayesian probability and manifold learning,the metric auto-encoder representation learning is analyzed.Experiments on public datasets such as mnist and cifar-10 show that compared to existing regularized auto-encoders algorithms,metric auto-encoder algorithm has better performance and higher recognition accuracy.Finally,for the feature fusion problem in image recognition,Gabor feature auto-encoder is proposed.The Gabor feature can be generally used to extract the texture features of the image.The image Gabor features and the auto-encoder algorithm are combined so that the representation learned from the encoder can reflect the image texture information.In order to extract features at different scales and locations,the paper uses a Gaussian pyramid and local sampling method.Experiments on the mnist public data set and license plate character samples show that the Gabor feature auto-encoder can capture the internal structure of the data and has higher recognition accuracy than the existing regularized auto-encoder algorithms.