Research And Application On Feature Learning Method Based On Deep Autoencoder Neural Network

Deep learning has become a research hotspot in artificial intelligence field at present.As a classic feature extraction algorithm in deep learning,Deep Autoencoder neural network(DAE)has been widely studied.The biggest advantage of DAE is that the data is extracted spontaneously layer by layer through a depth model composed of multiple hidden layers,and is widely used in image recognition?medical diagnosis?emotional analysis and other fields.This paper mainly studies the feature extraction method of DAE,and applies the extracted features to the field of recognition and classification.Thus the feature extraction ability of the DAE is measured by the corresponding accuracy.The research in this paper is mainly carried out from the following three aspects:(1)A deep self-adaptive balance autoencoder(DSBAE)is proposed for slow gradient transmission caused by internal covariate migration.Firstly,the parameters of each encoding layer are obtained through pre-training;Then the encoding layer and the balance layer are nested to form DSBAE.The parameters of encoding layers are balanced adaptively layer by layer using scaling and translation parameters in the balance layers.After that,the processed features are input into the nonlinear activation layer for mapping.Finally,the parameters both of the balance layer and the encoding layer are fine-tuned by gradient descent algorithm.DSBAE effectively solves effect of the internal covariate migration on feature learning.To verify the effectiveness of the proposed algorithm,DSBAE has been applied to handwritten digits recognition,and compared with DAE and other algorithms.The experimental results show that the classification accuracy of the proposed method is significantly improved.(2)Aiming at the problem of insufficient generalization and discriminating of features extracted of DAE,a deep semi-supervised autoencoder(LESDAE)with label and energy constraints is proposed.Firstly,it proposes energy items based on statistical physics,constrains the network to fit the distribution probability of sample data,and improves the generalization performance of the network.Meanwhile,according to the label constraint,label error is calculated for adjusting the network parameters to enhances the discriminability of features.To verify the effectiveness of the proposed algorithm,LESDAE has been applied to handwritten digits classifiaction,and compared with DAE and other algorithms.Through analysis of multiple experiments,LESDAE obtains better classification effects.(3)A variable learning speed deep autoencoder(VLSDAE)is proposed to deal with the problem of inefficient feature learning caused by fixed learning rate.The network firstly reflects the macro training effect based on the multi-scale error of batch,epoch and phase.Meanwhile,weight update correlation of each neuron is proposed to describe the micro training effect.The learning rate change coefficient is calculated via integrating the macroscopic and microscopic training effect description methods,and adaptively updating the learning rate,which improves the efficiency of feature extraction.This paper establishes a face recognition system based on VLSDAE.Through the application of face recognition,it is proved that VLSDAE extract more robust features,and achieve better recognition result.Through the classification on handwirtten digits,and compared with DAE and other algorithms,it is proved that VLSDAE can obtain higher classification accuracy under the condition of small training error.