Researchers develop technology to boost machine learning ability of AI

By Park Sae-jin Posted : September 10, 2019, 16:26 Updated : September 10, 2019, 16:32

[Gettyimages Bank]

SEOUL -- South Korean researchers have developed a computing technology that would greatly reduce the time required by artificial intelligence to analyze and sort data through a machine learning process.

Machine learning is an application of artificial intelligence (AI) that provides the ability to automatically learn and improve from experience. The self-studying process of AI mainly involves big chunks of raw data which are normally unlabeled and unstructured. Deep learning is a form of machine learning carried out by a neural computing system designed to mimic the thinking process of a human brain.

The Electronics and Telecommunications Research Institute (ETRI), a state-funded telecommunication technology research body, said in a statement that its AI research team has implemented high-speed processing optimized for distributed computing, a machine learning method which involves multiple computers to distribute data simultaneously to maximize the computing ability of an AI system.

Distributed computing is widely used in the analytics of big data and other information to reduce time but sometimes the bottleneck phenomenon occurs because of data traffic congestions, causing lag in computing performance.

According to ETRI, researchers have resolved the bottleneck phenomenon by creating a shared memory device called "Memory Box". The memory device works as a virtual memory system for a network of computers to accelerate the transfer of data between computers.

ETRI said that by adding Memory Box to an AI system, time taken in the deep learning process can be greatly reduced without replacing key devices. Through repeated learning, an AI system equipped with ETRI's new invention took 7 minutes and 31 seconds to sort 1,280,000 copies of photographic images into 1,000 categories, faster than the normal deep learning process without Memory Box which took 16 minutes and 23 seconds. Depending on the status of a deep learning framework, an AI's processing speed can be accelerated up to 4.2 times, ETRI said.
기사 이미지 확대 보기
닫기