An improved dropout method and its application into DBN-based handwriting recognition

As a typical deep learning method, Deep Belief Network (DBN) and Dropout method are usually used together for pattern recognition in case of lacking training data. Dropout training can avoid the overfitting phenomenon in deep neural network. During the testing stage, the outputs of all neurons in hidden layers are multiplied by a same factor as their actual outputs in the original Dropout method. It does not consider that the participation of a few poor-recognition models may reduce the recognition accuracy of its whole DBN model. This paper proposed an improved Dropout method, which can further increase the recognition accuracy of DBN by introducing probability statistics. The comparison of recognition results for MNIST handwritten digit database shows that the improved method can outperform the original and traditional identification methods under the same conditions.