In the whole brain architecture approach toward AGI, building a machine learning algorithm that corresponds to local circuit of the hippocampus is indispensable. However, extreme complexity of the hippocampus stands in the way. In this article, as a preliminary step toward that, we build a framework of the entorhinal cortex(EC), the gateway of information flow of the hippocampus. Based on neuroscientific knowledge, we define the interface and semantics of input to and output from the EC and specify the functions of the EC. Then, with this framework, we discuss the plausibility of existing computational hypotheses.

我々は階層型強化学習アーキテクチャRGoalを核として、ヒトの前頭前野周辺の情報処理機構のモデルを構築することを目指している。RGoalの行動価値関 数のテーブルを素朴に実装するとサイズが大きくなり、学習速度と汎化性能が悪くなるという問題がある。そこで、単一化の機構を用いてテーブルを圧縮する手法を提案する。また、テーブルが理想的に圧縮された場合に知識がどのような表現になるかを記述する記述実験を行うことで、提案手法の有効性の予備的検討を行う。RGoalは強化学習を用いたプログラム合成システムと見なせるが、合成の対象となる「プログラミング言語」の設計は、性能に大きく影響する。そこで記述実験で記述したテーブルをプログラムと見なし、その性質について考察する。

This study aims to design a virtual world agent's behavior based on the cognitive architecture ACT-R that aims to simulates the human mind. At this time, the hierarchical architecture is implemented to combine modules of decision making and body movement by using ACT-R and a three-dimensional game engine respectively. Each module collaborates to work via the communication server, realizing independency between two modules. The goal of this study is to develop a method of designing various behaviors of agent based on the hot swapping the ACT-R model or the modulation of parameters of the model.

Symbol emergence in robotics is regarded as a challenge to develop an artificial general intelligence that can adapt to real-world environment. In this talk, I will talk about the backgrounds and goals of symbol emergence in robotics and achievements in the field. I will also talk about fusion of deep learning and probabilistic generative models.

Imitation of behavior is one of the important elements of the development of intelligence, as can be seen from the fact that human babies grow up while imitating the behavior of the surrounding human beings. Such imitation learning of behavior was not easy because robots were expensive and algorithms were complicated. However, recently humanoid robots have become inexpensive, and the progress of deep learning has made algorithms development easier. Therefore, experiments were applied to imitation learning of behavior by humanoid robots, and the results are reported. The relationship between this method and the symbol grounding problem also be described.

We will investigate how science fiction (SF), which has drawn science and technology and the process of acceptance in the society in the form of a story, influences the development of artificial intelligence. The paper organizes and visualizes discuss the relationships between AI, SF and society with several examples. We want to propose a design method to the future society based on the recent possibilities and problems, and we hope to contribute to humankind about pioneering future technologies and societies.