A hybrid learning model based on auto-encoders
Ju Zhou, Li Ju, Xiaolong Zhang · 2017
The existing auto-encoder algorithm has been used to do deep learning. A variety of improved auto-encoder algorithms still have their disadvantages. In order to improve the learning accuracy of the auto-encoder algorithm, a hybrid learning model with a classifier is proposed. This model constructs a new depth auto-encoder model (SDCAE) by mixing a denoising auto-encoder (DAE) and a contractive auto-encoder (CAE). The weights are initialized by the construction method of the stacking auto-encoders, which is optimized by the gradient descent method. Therefore, the model has robustness to the reconstruction input of DAE and to the hidden layer representation of the CAE at the same time in the pretraining process. The learning model uses Softmax regression as a classification layer. The experimental results show that the classifier based on SDCAE has higher classification accuracy compared to existing auto-encoder on the given data sets.