A three-stage learning algorithm for deep multilayer perceptron with effective weight initialisation based on sparse auto-encoder

A three-stage learning algorithm for deep multilayer perceptron (DMLP) with effective weight initialisation based on sparse auto-encoder is proposed in this paper, which aims to overcome difficulties in training deep neural networks with limited training data in high-dimensional feature space. At the first stage, unsupervised learning is adopted using sparse auto-encoder to obtain the initial weights of the feature extraction layers of the DMLP. At the second stage, error back-propagation is used to train the DMLP by fixing the weights obtained at the first stage for its feature extraction layers. At the third stage, all the weights of the DMLP obtained at the second stage are refined by error back-propagation. Network structures and values of learning parameters are determined through cross-validation, and test datasets unseen in the cross-validation are used to evaluate the performance of the DMLP trained using the three-stage learning algorithm. Experimental results show that the proposed method is effective in combating overfitting in training deep neural networks.