A Hybrid Depth Network Learning Model Based On Auto-encoders

In recent years,deep neural network is a hotspot in the field of artificial intelligence and big data analysis.The depth neural network discovers the distributed feature representation of the data by combining the low-level features to form more abstract high-level representation attributes.At the same time,auto-encoders are constructed as the basic framework of the depth neural network,through the unsupervised feature learning process to extract data features or attributes,it has made significant achievements in prediction and identification of machine learning.In order to make better use of the advantages of auto-encoder model on feature learning,this thesis designs a new hybrid depth learning model based on traditional auto-encoders,which incorporates the training process and the constraint condition of Denoising Auto-encoder(DAE)and Contractive Auto-encoder(CAE).The DAE makes it possible for the auto-encoder to reconstruct the input better in the training learning process by adding a certain proportion of noise interference to the input data,so that it can obtain a characteristic expression with robustness to noise interference;The regularized target in the CAE's loss function is the Frobenius norm of the Jacobian matrix of the encoder,which reduces the influence of the minimizer on the encoder and to assist the encoder with better learning characteristics.The hybrid learning model proposed in this thesis is to build the depth structure by means of module assembling.The DAE unit and the CAE unit are introduced into the same neural network to form a new depth learning model SDCAE.Which makes full use of the advantages of two auto-encoders in the training process and uses the gradient descent method on training the parameters of the classifier and the auto-encoder in the model.This thesis validates the model on the experiment with MNIST dataset.By comparing the classification experiments on three kinds of automatic encoders,the learning accuracy in the MNIST has been improved on the new hybrid learning model SDCAE,which shows better learning performance.