The neural network developed by Yoshihiro Yamada (1st year doctoral student), a graduate student in the Intelligent Media Processing Laboratory, Department of Electrical and Information Engineering, Graduate School of Engineering, Osaka Prefecture University, has achieved the world's highest recognition accuracy in the field of general object recognition. ..In December 2016, Mr. Yamada proposed a method with the highest recognition accuracy in the world at that time. In May 12, the newly proposed neural network once took its place, but this time, we developed a neural network with a significantly improved recognition rate compared to the conventional method, and once again achieved the world's highest recognition accuracy. Achieved.
General object recognition is a task (processing execution unit) that recognizes (classifies) various objects such as "airplanes", "cars", "birds", and "cats".In recent years, methods using neural networks, which have been attracting attention in deep learning, have become the mainstream, and fierce R & D competition has unfolded in which records are set 2 times in the past two years with only the main ones. Has been done.
The method called Shake-Shake, which broke the record of Mr. Yamada's proposed method in May 2017, was a new method that showed that higher performance can be achieved by "moderately interfering with learning", but it is relatively shallow. It had the drawback of being applicable only to networks.Therefore, Mr. Yamada realized Shake-Shake's learning method that "moderately interferes with learning" in a form that can be applied to networks with deep structures.In the process, we also introduced a device to stabilize learning and applied it to the method that we proposed last time and achieved the highest recognition accuracy in the world.As a result, we succeeded in significantly improving the recognition rate and achieved a recognition accuracy of 5%, which is a large difference of about 3% from the method that was the best in the world so far.
This result is expected to have a great ripple effect not only on general object recognition tasks but also on various image-related tasks.
