Exploring the essence of intelligence
The Dwango Artificial Intelligence Laboratory was founded in 2014 with the aim of “creating artificial intelligence as a gift for the next generation” that will contribute to the sustainable development of society and pioneer human intellectual frontiers. For this purpose, we are concentrating on fundamental research aimed at realizing the highly versatile artificial intellectual ability of creative problem solving, which is not possible by current artificial intelligence technology.
Recently, machine learning technology such as deep learning has progressed to the point where given sufficient data, a “narrow” AI with abilities equal or even superior to humans can be constructed for specific problem domains. It is possible to build a “big switch statement” type AI that bundles various narrow AI, but such AI cannot create new solutions under unfamiliar situations where data is inadequate.
Against this backdrop, technology that can generate solutions in unfamiliar situations is essential for building human-level intelligence. However, humans cannot necessarily understand the problem solving capabilities that machine learning exerts. Furthermore, it is not obvious that the information processing mechanism of the brain acquired by evolution is understandable by humans. Therefore, we are also betting on the possibility that such intelligence can be realized without fully understanding the essence of creative problem-solving capability.
At Dwango Artificial Intelligence Laboratory, we are employing both of the following approaches in order to explore the essence of such intelligence
1) Theoretical Approach Based on an Understanding of General Intelligence:
This approach aims to understand the intrinsic properties of intelligence by elucidating the information and calculation process, or by further reducing these components into more basic elements. Research and development of intelligent technology attempts to realize reasoning in unknown situations by flexibly combining various pieces of information and diverse concepts acquired mainly by machine learning. In particular, it is important to reformulate the problem so that the unknown situation can be regarded as a known situation. This is accomplished through a process of using technologies to decompose knowledge to make it easy to associate with symbols, methods that allow for mapping knowledge to different areas, and techniques to recognize, control and learn from a meta-level in order to combine knowledge. Currently we are conducting research on action abstraction, equivalence structure extraction, and so on.
2) The whole-brain architecture is an approach referring to the natural general purpose intelligence: the Brain.
Based on the premise that it is not always easy to understand the essence of intelligence, this approach is aimed to build a general purpose intelligence by referring to an architecture that can flexibly formulate a problem, which is thought to be possessed by the brain. This approach was defined “‘to create (engineer) a human-like artificial general intelligence (AGI) by learning from the architecture of the entire brain” at the NPO corporate whole brain architecture initiative (WBAI). Currently, we support the development of open platforms directed by the WBAI such as connectome informatics and middleware for brain-inspired AI. Our lab will gradually expand R&D of WBA on these platforms. In order to promote such R&D, our laboratories will need to coordinate a wide range of knowledge and skills including artificial intelligence, machine learning, neuroscience and cognitive science.