AlphaGo algorithm applied on mobile devices in future

KAIST presented article about DRL AI semiconductor technology / May be utilized in intelligent robot driving automatic drone and game

News provided by wipnews
2021-08-10 00:29:26 KST language
[Photo provided = MSIT]

[Photo provided = MSIT]

Ministry of Science and ICT (MSIT) announced that research team of Professor Yoo Hoi-jun of Korea Advanced Institute of Science and Technology (KAIST) has developed a semiconductor technology that can process Deep Reinforcement Learning (DRL) with high performance and power efficiency. DRL was utilized in ‘AlphaGo’ a go AI program developed by Google DeepMind.

The research was nominated as the highlight article among 200 articles presented in Institute of Electrical and Electronics Engineers Very Large Scale Integrated Circuit Symposia.

DRL algorithm uses multiple nerve networks to find answers. The networks have complicated structure and has to process large-scale data so it may be only realized by parallel arrangement of high-performance computers and laptop and smartphones could not realize the algorithm due to its limited arithmetic capacity and memory.

The research team developed the ‘OmniDRL’ a semiconductor technology that has higher performance and 2.4 times efficient power that may realize DRL on mobile devices.

Specifically they utilized Static RAM based Processing-In-Memory (PIM) semiconductor technology that increases compressibility of deep neural network data enables processing in compressed data and integrates processing and storage function. Particularly conventional PIM semiconductors could only process integer level data but the research expanded it to floating-point level.

‘OmniDRL’ was applied to ‘human robot adaptive walking system’ which is usually used in performance comparison of DRL algorithms and the adaptive walking was 7 times faster.

Professor Yoo said that the research enabled inference and learning under high data compression on single semiconductor and floating-point level processing. He expects that the technology could be applied in various fields like intelligent robot driving automatic drone games etc.

Song Kyung-hee MSIT director general of Artificial Intelligence Policy Division said that the research was internationally appreciated in the AI semiconductor a new paradigm in semiconductor area. MSIT is broadening its support to semiconductor industry by going to continue the trillion won support in R&D of AI semiconductor that has begun last year and starting 400 billion won project of PIM semiconductor development.

Website: : http://cms.wip-news.com/news/articleView.html?idxno=7930