This paper presents a compression method to reduce the number of parameters in convolutional neural networks (CNNs). Although neural networks have an excellent recognition performance in computer vision application, there is a need for a large memory for storing amount of parameters and also necessary in a high-speed computational block. Therefore we propose two the compression schemes (pruning, weight sharing) in LeNet network model using MNIST dataset. The proposed schemes reduced the number of parameters of LeNet from 430,500 to 32 excluding index buffer size.
展开▼
