Reduce communication traffic in the same environment to increase AI learning speed
Even beginner developers can easily create AI models
| Development of a Deep Learning Dashboard for Domestic Developers

Deep learning is a technology that enables computers to think and learn like humans. However, it also takes a lot of time for computers to learn from large-scale data or models, such as video, images, and audio.

To process large-scale data, distributed learning techniques are used to reduce training time by utilizing multiple computers. However, even distributed learning techniques had limitations, such as communication bottlenecks occurring when running large-scale models simultaneously on multiple computers.

One way to solve this is to increase the performance of the CPU or GPU, which are responsible for the computer's computational and processing capabilities. However, this method entails a significant cost burden, such as the need to continuously upgrade equipment.

Domestic researchers have succeeded in developing computing technology that reduces AI training time by a quarter.

The Electronics and Telecommunications Research Institute (ETRI) announced on the 10th that it has developed high-speed processing technology optimized for distributed deep learning. By applying this technology, an AI model that used to take a week to train can be trained in 1 to 2 days in the same environment.

ETRI developed a shared memory device called the "Memory Box" to reduce training time by resolving communication bottlenecks that occur during distributed learning. The Memory Box is positioned between computers to facilitate the sharing of learned data among them, thereby reducing the amount of data transmitted.

Therefore, deep learning training time can be significantly reduced in the same environment with minimal investment without major equipment replacement. In particular, since it can provide both hardware and software forms, customized technology transfer is possible for the user.

As a result of conducting an experiment to train a model that classifies 1.28 million images of 1,000 different types using memory boxes 10,000 times, the existing server method took 16 minutes and 23 seconds, whereas the method using this technology took 7 minutes and 31 seconds.


Development of a deep learning dashboard providing a dedicated GUI
Global IT companies such as Amazon, Google, and Microsoft have attracted developers and increased their market share in the AI computing infrastructure market by releasing their source code or providing large-scale computing resources in the form of the cloud.

As a result, domestic companies and institutions researching AI had to rely on the services of foreign companies or incur significant costs to build their own servers.

Accordingly, ETRI researchers also developed a 'Deep Learning Dashboard' to provide an AI computing environment where domestic developers can easily conduct deep learning research.

The dashboard developed by the research team provides a GUI, so developers do not need to enter code line by line. This helps reduce not only training time but also model development time. It also supports TensorFlow and Caffe frameworks, which are commonly used in AI development, allowing you to train graphic models developed on the dashboard.

This technology can be utilized in various industrial fields requiring deep learning and AI, such as high-resolution medical image analysis or massive image analysis. In particular, if used by small and medium-sized enterprises, schools, and startups with limited computing resources, it can save development time and costs.

Currently, two small and medium-sized enterprises (SMEs) have received the memory box technology and are pursuing the establishment of research institute companies. The research team aims for commercialization next year through these companies.

Choi Wan, a principal researcher at ETRI’s AI Research Institute and the project leader, stated, "We hope this will help replace the AI computing infrastructure market, which is currently dominated by global companies, with our own technology, and contribute to the development of high-difficulty deep learning technology and proprietary AI supercomputing systems."