Convolutional Neural Network And Its Applications

Convolutional Neural Network (CNN) is an efficient learning neural network. This paper provides detailed analysis about its basic theory and applications in image classification and large-scale continuous speech recognition.CNN combines feature extraction and classification process for neural network training and gets significant success in the field of image classification. In experimental section of this paper, MNIST handwritten digit recognition using CNN model gets better result than traditional methods.Since2010, Deep Neural Network (DNN) makes speech recognition word error rate reduce20%-30%, which is a great breakthrough in the field. In recent years, Academic research suggests that CNN can gain more performance improvement than DNN. The experimental results show that CNN-HMM hybrid model can continue to reducing the word error rate, compared to DNN-HMM hybrid model.The optimization process of CNN and DNN contains a lot of matrix operations. In large-scale image classification and speech recognition, the process will be very slow. Using the NVIDIA GPU to speed up the matrix operations can get dozens or even hundred times faster than CPU.In order to overcome the fitting problem, this paper also explores new methods like the Rectified Linear Units(ReLU) and Dropout. The experiments on image classification indicate ReLU and Dropout can improve the generalization ability of CNN.