• home_icon HOME
  • Research
  • Highlights
Research on control algorithms that can learn like a human

Research on reinforcement learning has been conducted in the engineering field as well as studied continuously for the past 20 years in decision making and the computational neuroscience field. Since the first research results were published in the late 1990s showing that the data processing procedure in the midbrain’s dopamine system involved in human and animal behaviors and the learning process could be explained by reinforcement learning theory, there have been many developments in reinforcement learning in the field of neuroscience.

The research team of Professor Sang Wan Lee is conducting artificial intelligence-neuroscience convergence research to understand the learning control mechanism of the prefrontal cortex and apply it to the design of an artificial intelligence system. This research has sometimes led to the development of various types of algorithms, but furthermore, it is meaningful because neuroscience-based reinforcement learning theory presents the possibility of solving various difficult engineering problems such as a design that balances performance, efficiency and speed. The present study proposed a “prefrontal meta-controller” theory based on the fact that human brains easily solve engineering problems that individual algorithms such as reinforcement learning cannot solve. The point of this theory is that the reward prediction signal or state prediction signal that can assess the reliability of learning about the external environment by itself is processed in the basal ganglia-ventrolateral prefrontal cortex network, and the human brain finds the learning and inference strategy most suitable to the changing external environment through a dynamic process that competitively and cooperatively integrates the information.

Studies on its applicability to the robotics and neuroscience fields based on brain-inspired artificial intelligence technology were published in papers in Science Robotics and in Current Opinions in Behavioral Sciences.

Publication of a paper unprecedented domestically and internationally

The approach in the present study is novel because unlike the paradigm of existing AI research that replaces a human’s work using AI technology, it combines the high-level learning and inference processes of humans with an AI algorithm to improve the learning and inference control ability of humans. If advanced functions of the human brain can be reflected in the design of an artificial intelligence system, its performance can be significantly improved, and on the other hand, if the quasi-optimal mechanism of humans is reflected in the design of an artificial intelligence system, a customized service based on a deeper understanding of humans can be provided. It is expected that the human and AI systems can interact at a deeper level in the near future.

Continuity of research using brain-inspired artificial intelligence technology

In the future, research will be continued including studies to design an artificial intelligence system based on a more fundamental principle or theory of brain function, beyond artificial intelligence research that transplants some computational modules of the brain into an algorithm, and research to verify complex brain science hypotheses that cannot be solved by conventional approaches of brain science by virtue of the latest artificial intelligence techniques. Professor Lee thanked the many people who gave him an opportunity, even though the research was a huge challenge, and explained that he would publish various research results based on the theory. In 2014, Professor Lee presented neuroscientific evidence for the first time in the academic world that the prefrontal-basal ganglia brain network circuit in humans can control heterogeneous reinforcement learning strategies, and in 2015 presented research that the same brain circuit controls the one-shot inference process through interaction with the hippocampus.

Prof. Sang Wan Lee
2018 KI Annual Report

KAIST 291 Daehak-ro, Yuseong-gu, Daejeon (34141)
T : +82-42-350-2381~2384
F : +82-42-350-2080
Copyright (C) 2015. KAIST Institute