This paper presents Artificial Brain Operating System which imitates the behavior of human brain and its algorithm mimics thinking process. I describe goals of the proposed system, which are a) voluntary and self-directed action of speech and motion, b) visual and auditory understanding, c) skill and knowledge acquisition based on image processing, d) internal information processing standardization.

エージェントに快と不快の情報を導入することで、強化学習分野における汎用性の獲得を目指している。現在の強化学習では報酬は一般的にエンジニアが定義する情報であり、エージェントは設定された報酬に対して獲得に繋がった行動を価値として算出し、 次回の試行時の行動選択に役立てる。しかし強化学習は特化型人工知能と呼ばれるように転移学習が困難という性質がある。本モデルでは、報酬としての快という情報を不快の解消として位置付けることで報酬情報を不快を発生させた事象と紐付け、エージェントに自律的に何が報酬(不快を解消させるのか)なのかという情報を獲得させるとともに、報酬情報を区別する能力を獲得させることで複数のタスクを同一のエージェントによって学習可能な状態にすることを目標としている。

This paper discusses the neural network applying the cell differentiation model. It is assumed that a largescale neural network will be required to deal with various applications flexibly in the future. This neural network aims to enable to inherit learning achievements, expand and improve its structure and parameters automatically.

多くの機械学習手法は，特定の問題に特化した形てでアルゴリズムを作成し学習を行うた め，他の問題が与えられたときには初めから学習を行う必要がある.そのような手法では多数の問題 に適応的に対処できる汎用的なシステムを作成することは困難である.一方で近年では転移学習や マルチタスク学習のような複数の問題を扱う手法が積極的に研究されるようになってきた.本発表で は転移学習やマルチタスク学習の枠組みを説明し，それらに基づいた汎用的なシステムを構築する ための現在の困難と今後の可能性について紹介する.

We postulate that Artificial General Intelligence remains elusive because of numerous undisputed assumptions that are deeply rooted into the traditional understanding of intelligence. We claim that these assumptions shape an anthropocentric bias that prevents the development of a general theory of intelligence capable of explaining the behavior of not only human and machine intelligence, but also any other entity that exhibits intelligent behavior. The most important of these assumptions is the failure to recognize darwinian evolution as an intelligent entity despite the growing consensus about its superior capabilities to develop biological contrivances. In order to avoid underrating and neglecting evolution as intelligent, other assumptions must be dropped. Such is the case for the requirement of language, which is only relevant in social contexts. Moreover, the boundary of evolution as an agent distinguished from the environment is not well-defined, which suggests that agent boundaries are redundant in General Intelligence and results in an equal treatment of polymorphic robots and multi-agents, to name a few. By revealing these and other assumptions, we propose that human intelligence should be relieved from standing at the center of studies about General Intelligence.

We propose a framework of the study of computable induction, which is a computable version of Solomonoff's universal induction. As a concrete example, we consider a problem of confirmation, such as the probability that the sun will rise tomorrow. Although we can not tell the exact probabilities, we can deduce the rate of the convergence up to a multiplicative constant, which is slightly faster than Laplace's result.

On the symbol grounding problem, it is better to consider separating symbols for internal processing of thinking and symbols for communication. And it is essential that symbols for communication point to shared believes. Therefore, this paper reports of experiments to generate shared believes on the symbol grounding problem using the deep generation model.

In our paper we discuss the problem of tacit knowledge which probably is one of the biggest obstacles on the way to human-level language understanding. While the latest massive transformer-based NLP algorithms show the potential to generate natural text, translate and answer questions, these achievements are still insufficient to directly help machines to acquire common sense. We introduce the problem of tacit knowledge and various approaches to solving it by knowledge completion with automatic text generation which is meant to enrich existing texts and to improve machine learning process by filling the semantic gaps naturally omitted by human beings while generating a natural language. We present various types of knowledge additions we performed in the past and report on the latest achievements in Schankian scripts generation.