Professor Tatsuya Harada, Graduate Student Masatoshi Hidaka, Graduate Student Yuichiro Kikura, and Lecturer Yoshitaka Ushiku of the Department of Intelligent Machinery and Information Science, Graduate School of Information Science and Engineering, the University of Tokyo, are on the Web browser that is installed as standard on commercial PCs and smartphones. Announced that it has developed a software framework "WebDNN" that can execute deep neural networks (DNN) at high speed.
DNN is a form of artificial intelligence that has been rapidly developing in recent years, and is said to be an effective method for recognizing and generating images and sounds.However, the calculation load is high, and it is difficult for the service provider to easily try it because it is necessary for the service provider to prepare a large number of computers according to the number of users or install a dedicated application on the user's terminal in order to incorporate it into the Web service. rice field.
As a means to solve these problems, the idea of incorporating software that processes DNN in a Web page has been proposed for some time, and this research has realized this. We have developed a technology that eliminates redundant calculations in the DNN and a technology that maximizes the capabilities of the terminal by utilizing the latest standards of the Web, and provided it as a system that can be easily incorporated into Web services.
With Web services equipped with WebDNN, users can simply open it in a Web browser and execute DNN at high speed on the terminal.In the experiment, the execution speed of the image recognition process was about 50 times faster than the conventional one.
This system is useful as a basis for anyone to easily try out the latest results of artificial intelligence research such as image generation, and is expected to be applied to various demonstrations and applications.
Paper information:[Proceedings of the 2017 ACM on Multimedia Conference] WebDNN: Fastest DNN Execution Framework on Web Browser
