A study on accelerating convolutional neural networks

Recent deep-learning methods have been paid more attention than shallow-learning ones because they have deep and complex structures to approximate functions. The salient feature of deep neural networks is to use many layers where many of them are used to extract data features and few are for classification or regression. The most severe problem of a deep neural network is using too many parameters that cause too much memory usage and computing resources for both training and inference. Thus, deep learning approaches are not suitable for real-time industrial applications that have limited computing resources such as memory and CPU. For example, a famous convolutional neural network (CNN), AlexNet, uses up to 60 million parameters to train ImageNet dataset and many imaging projects apply AlexNet to their own applications as transfer learning. Thus, this work proposes a feasible solution to trim the CNN, speed it up, and keep the accuracy rate similar. Two main types of CNNs and AlexNet, were validated, respectively, in THUR15K, Caltech-101, Caltech-256, and GHIM10k datasets. The results show that the parameter amount greatly decreased (76%) but the recognition rate dropped slightly (1.34%).