A CUDA-based implementation of convolutional neural network

Training of the convolutional neural network (CNN) entails many iterative computations. In addition, as a depth of neural network has increased and the number of training data has become large in recent years, the amount of computation required for the training has also dramatically increased. The parallel operation using the hardware feature of general-purpose computing on graphics processing units (GPGPU) has been known to be suitable to be applied to neural networks. Accordingly, this paper proposed a method to improve a training time by parallelizing a training algorithm in the CNN using the Compute Unified Device Architecture (CUDA) capable of controlling NVIDIA GPGPU. The experiment result showed that the proposed method in this study improved about 2.5 times faster training speed than existing training method that makes use of only a single CPU.