Research On Representation Learning Algorithm Based On Auto-encoder

Machine learning,as a method to realize artificial intelligence,has become one of the hotspots of present research.As the basis of machine learning,the importance of representation to algorithms is unquestionable.In other words,the performance of machine learning algorithms depends to a large extent on the choice of representation.Representation learning is able to adaptively learn the effective representation of data according to the practical application,and the performance is good.Auto-encoder,as an important representation learning algorithm,has attracted many researchers' attention.In this thesis,the supervised metric auto-encoder and Gabor auto-encoder are proposed based on the theory of metric learning,Siamese network,Gabor filter and auto-encoder.The main contents of this thesis are as follows:First of all,the thesis provides a theoretical foundation for the following chapters.Considering that the new algorithm uses the theory of auto-encoder for reference,this thesis mainly introduces the basic theory of artificial neural network,back propagation algorithm,auto-encoder and its improved algorithm,and describes several datasets used in the experiment.Secondly,this thesis proposes a supervised metric auto-encoder based on the ideas of metric learning,Siamese network and auto-encoder,so that the learned representation has the characteristics of reconstruction and metric learning at the same time.In order to explain and understand the algorithm,this thesis gives a detailed derivation process,algorithm flow and network architecture.Through the experiments on MNIST and Fashion-MNIST datasets and the comparison with other algorithms,this thesis proves that the algorithm represents better learning performance and higher recognition rate.Finally,based on the theory of Gabor filter,Siamese network and auto-encoder,this thesis proposes a new algorithm: Gabor auto-encoder.In theory,the combination of auto-encoder and Gabor filter will make the representation of Gabor auto-encoder learning not only have the reconstruction characteristics of auto-encoder,but also the image texture information.Similarly,this thesis describes the detailed derivation process,algorithm flow,network architecture and image recognition process of the Gabor auto-encoder.The experiments on MNIST,Fashion-MNIST,MNIST-rot and license plate datasets show that the Gabor auto-encoder performs extremely better than traditional machine learning algorithms and some convolutional neural network algorithms.