Though it may come as a surprise, North Korea has been pursuing the development of artificial intelligence technologies since the 1990s through its series of state-run tech software research agencies.

Once recognized as a relative leader in AI computing, North Korea’s AI development is expected to hit hurdles due to strained resources, suggests a new report released Tuesday by KDB Research.

The economic think tank under the state-run Korea Development Bank said that North Korea’s AI development will remain limited as the country finds it increasingly difficult to invest in the high-tier computing hardware needed to train and improve AI systems. 

“Modern AI technology evolves by collecting and feeding large amounts of data to a high-power computing system. This requires enormous hardware and software investment,” said the report’s author Kim Min-kwan from KDB Research’s unification business unit.

“However, North Korea’s AI development is expected to hit a wall, due to its strained financial resources, economic situation and international sanctions including the Wassenaar Agreement, which prohibits weapons export to countries associated with terrorism,” Kim said.

North Korea has been developing AI technologies from the 90s via the state-run software research institute Korea Computer Center, the report noted. The country has three other research centers each dedicated to AI development, manufacturing automation as well as the development of digital security and biometric authentication software, it said.

A hallmark of North Korea’s AI development program is Eunbyul, its self-developed AI designed for the ancient Chinese game of Go.

Before the emergence of Google DeepMind’s famed AlphaGo, Eunbyul, developed by the Korea Computer Center in 1997 had been the reigning champion of international computerized Go competitions. It won the championship six times -- in 1998, 2003, 2004, 2005, 2006 and 2009.

The core technology underpinning Eunbyul is in fact no different from that of AlphaGo. Both operate on the Monte Carlo tree search method, a heuristic search algorithm which decides the most promising move based on knowledge previously learned via machine learning.

Yet, the major difference is that while Eunbyul’s 2010 edition was trained on just 16 CPUs, AlphaGo during its match against Go grandmaster Lee Se-dol in 2016 used 1,920 CPUs and 280 GPUs to consider 100,000 probable outcomes per second, the report noted.

North Korea is also known to possess deep learning capabilities, a form of high-level machine learning which uses large artificial neural networks — layers of interconnected nodes — that rearrange themselves as new information comes in, allowing them to self-learn without the need for human programming.

Recently, the country managed to vastly improve its voice recognition software Ryongnamsan using deep learning algorithms, according to a 2017 paper in Pyongyang-based Kim Il-sung University’s scholarly journal, cited by KDB Research.

It is the same tech powering AI voice assistants like Apple’s Siri, Amazon’s Alexa and Google’s Assistant, the KDB report said.

In addition, North Korea is reportedly developing AI systems for application to facial recognition and finger recognition software. It also is said to possess knowledge of complex computing methods such as the vector space model and support vector machine, KDB said.

Despite possessing fundamental AI technologies, North Korea will increasingly find it difficult to move forward its data-driven AI development, due to strained resources stemming from its international isolation, the report concluded. 

By Sohn Ji-young (jys@heraldcorp.com)