One Step Closer to Flexible AI

A KAIST team has found a way to solve overfitting and underfitting in machine learning.

The Korea Advanced Institute of Science and Technology (KAIST) announced on Jan. 5 that professor Lee Sang-wan and his research team at its Department of Bio and Brain Engineering found a way to solve overfitting and underfitting in machine learning.

Artificial intelligence models present optimal solutions to various problems. However, their situational flexibility is still at a very low level. On the other hand, people focus on given problems while responding flexibly to changing conditions and situations.

The research team prepared a theoretical basis on how the human brain does so and derived a meta reinforcement learning model using brain data, stochastic process inference, and reinforcement learning algorithms. According to its theory and meta reinforcement learning model, various conflicting situations in the real world can be solved by artificial intelligence and the human flexibility can be indirectly measured using simple games.

According to the team, its discoveries can be utilized in next-generation artificial intelligence development, smart learning, cognitive behavioral therapy, and so on. Details of the research are available in the online edition of the Cell Reports journal.