A Deep Learning Method Based on Hybrid Auto-Encoder Model

In order to make Auto-Encoder improve the ability of feature learning in training, reduce dimensionality and extract advanced features of more abstract features from mass original data, it can improve the classification results ultimately. The paper proposes a deep learning method based on hybrid Auto-Encoder model, the method is that CAE (Contractive Auto-Encoder) and traditional neural network SRBM (Sparse Restricted Boltzmann Machine) are combined to form a hybrid Auto-Encoder network to train dataset. This model is constructed to achieve a variety of neural network between functionally complementary advantages and strengthening reducing dimensionality, extracting more "good" feature and enhancing the learning ability of the hybrid model. In order to speed up the convergence rate of hybrid model parameters, when training update parameters, we add a method of updating parameters which is Polyak Averaging. The experimental results show that the hybrid model can extract more accurate features and improve the accuracy of classification. Polyak Averaging accelerates the speed of convergence of the model. It is demonstrated a comparison of the classification accuracy with the single stacked two layer CAE training model experimentally, the classification accuracy of the model is proved to be better.