Kim Sang-wook, a professor of Hanyang University Department of Computer Science, and his team recently developed a new distribution deep learning technology, “ALADDIN,” which has twice the function of previous technologies, according to Hanyang University on October 27.

Deep learning is a technology of an enormous amount of big data learning using a model of numerous layers and is considered as a core technology in the 4th Industrial Revolution era. However, deep learning for big data learning required a vast amount of time and resources, which was a huge obstacle in the learning research. Based on this, active research in academics and industries has been done to speed up the development of deep learning technology.

ALADDIN is a distribution deep learning technology that accelerates deep learning based on distributed clusters comprised of tens and hundreds of workers. The gist of ALADDIN is investigating the cause of deep learning function degradation through detailed analysis on previous distributed deep learning methods and solving them.

Data processed by the workers are used for updating a global model in the parameter server. Each worker sends the processed data to the parameter server and receives the global model that reflects such data. Thus, the worker and the parameter perform not an asymmetrical communication where one side sends the data unilaterally but asymmetrical communication that conducts both sending and receiving.

Professor Kim’s research team found that the symmetrical communication between workers and parameter servers was the fundamental cause of the function degradation and designed a new distribution deep learning technology based on asymmetrical communication. They applied the strategies to solve the accuracy function issues raised by asymmetrical communication as well. As a result, ALADDIN was able to maintain the accuracy, unlike the previous distribution deep learning methods, while enhancing the performance by 200%.

Professor Kim’s ALADDIN research holds significance in that it can apply to future technologies as well as already existing deep learning technologies. Therefore, ALADDIN is evaluated to be having a high potential to be used in diverse fields of AI.
Professor Kim’s ALADDIN research was supported by BrainKorea21, the National Research Foundation of Korea, and the Institute for Information & communication Technology Planning & evaluation. Also, Hanyang University Professor Ko Yun-yong, researcher Choi Ki-bong, SK Telecom researcher Je Hyun-seung, and the United States Pennsylvania State University Professor Lee Dong-won joined the research.

ALADDIN was highly recognized for its technology originality and excellence and is planned to be presented in the “The ACM International Conference on Information and Knowledge Management (ACM CIKM) 2021” held for five days starting from November 1. The ACM CIKM is one of the conferences that is acknowledged in the data science field.