人工知能学会第二種研究会資料
Online ISSN : 2436-5556
2019 巻, AGI-013 号
第13回汎用人工知能研究会
選択された号の論文の11件中1~11を表示しています
  • 深田 博幸
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 01-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    現在のディープラーニングなどのAIシステムは、予め膨大なデータを使って学習するものが一般的である。しかし人間や生物の赤ちゃんは、予め膨大なデータを学習してからのみ動く訳ではない。また事前に学習が必要なシステムは、お決まりの答えしか出せない。然し、人間や生物は状況に応じて、柔軟な対応ができている。そこで逐次に事象を学習し、事象の価値を評価し、有意性を見出した場合には、新たな行動を起こし、問題の発見、解決、発案や強化学習などを行うことを可能とする方式を考案した。また人間の感性をモデルとして、多様な情報に対し、同様な評価を可能とする方式を考案した。

  • 江村 憲夫
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 02-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    For external information of any scene and activated internal information, when the Intention Active Information and associative/hierarchy memory are activated at the same time, based on the activated information, the problem based on the intention of the intention and the solution means (will of the artificial brain) generate a concrete execution means. This ability to think is a supplement of missing information, control of the abstractness of information, association / hierarchical connection of each information and enhancement of activation propagation, etc., along with making full use of associative / hierarchical memory. This ability also depends on the formation of a hierarchical spread of Intention Active Information with an image (problem consciousness, desire, sense of crisis, etc.). In this paper, I introduce current initiatives and problems to be solved.

  • 岡谷 基弘
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 03-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    How to make the AGI (Artificial General Intelligence)" is a fundamental and difficult problem. The AGI may be made referring to human intelligence. I introduce a study of human intelligence as functions. There are three functions in human intelligence. The first is "the symbolization of the world". The second is "the desire which make us aim to realize the desirable world". The third is "the decision system based on prediction". This way of thinking allows us to consider the AGI design as approximation of human intelligence.

  • 岩崎 博生, 伊野 文彦
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 04-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    本論文では,推論モデルにおけるメモリ使用量の削減を目的として,深層ニューラルネットワークに対する圧縮手法を提案する.提案手法は,ネットワークを構成するユニットを統合するために,非構造的枝刈りをもとに,重み行列に対して k平均クラスタリングを適用し,各行を k個のクラスタに分類する.さらに,各クラスタに属する行を,クラスタの中心,すなわちクラスタに属する行の平均値に置換する.置換結果に基づいて,ネットワークのユニットを統合することで,重み行列の列数および行数をそれぞれ k個に削減し,メモリ使用量を削減する. kを変えながら提案手法を に適用したところ,適用前の予測精度からの精度低下をたかだか 程度に抑えながら,全結合層のメモリ使用量を最大で 削減できた.また,同一メモリ使用量削減率における提案手法と非構造的枝刈りの予測精度を比較したところ,提案手法の予測精度が非構造的枝刈りの精度よりも最大で 高いという結果を得た.同様に,同一予測精度における提案手法と非構造的枝刈りのメモリ使用量削減率を比較したところ,提案手法のメモリ使用量削減率が非構造的枝刈りの削減率よりも最大で約 大きいという結果を得た.

  • 疋田 聡
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 05-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    In this study, "AGI" is used in the sense of creating AI as general-purpose as possible, and consciousness is not essential. Existence of consciousness is a viewpoint of "strong AI" and "weak AI", it is different from generality. Nevertheless, since consciousness was created in the process of evolution, having consciousness is considered advantageous for survival, and it may be used to improve AI performance. Therefore, we conducted experiments using deep generation models to improve the performance of AI with several functions in consciousness.

  • 中野 敬三
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 06-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    Society 5.0, proposed the government's 5th Science and Technology Basic Plan, is realized by a system that highly integrates cyber space and physical space. An Important factor of cyber space is autonomous AGI. Knowledge Accumulated By Mankind is a social asset, and autonomous created from this knowledge salsa social asset. Autonomous AGI has a social adult ego and works for a human being's well-being and must accept human common values. A new society of Society 5.0 solves social problems and realizes economic prosperity. Autonomous AGI with the estate of social adults solve the social problems of war and environment which are caused by human selfish desires. Autonomous AGI productivity can be considered social productivity. Society 5.0 will be a mixed economy society where a communist economy based on autonomous AGI productivity and a capitalist economy based on human activity coexist. Products Produced by autonomous GI can be distributed equally to the publics social income. This social income not only reduces the economic disparity of the people and fosters a sharing society, but also increases effective demand needed for economic development.

  • 長島 一真, 森田 純哉, 竹内 勇剛
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 07-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    A development of autonomous agents requires not only learning by external rewards but also a mechanism for generating spontaneous rewards from the inside. Specifically, reinforcement learning methods using intrinsic motivation have been studied. In this study, we propose a method to express it by ACT-R, a general-purpose cognitive architecture. The proposed method focuses on a compilation of pattern matching in ACT-R and regards pattern discovery as a source of intrinsic motivation.

  • 田和辻 可昌, 中村 遥佳, 松居 辰則
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 08-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    In the field of cognitive neuroscience, various cognitive architectures have been proposed to provid the bridge between the cognitive function and neural structure. However, these architectures are highly individual and the models based on different architectures are not likely to be referenced among researchers. In this study, integrative description for organizing viewpoints for these models is proposed, in which the devices are defined as cell populations which transfers a kind of information into another. As a domain description, the devices subserving saccadic eye movement are defined based on the device ontology.

  • Masao Yokota, Rojanee Khummongkol
    原稿種別: SIG paper
    2019 年 2019 巻 AGI-013 号 p. 09-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    Spatiotemporal (or 4D) representation and reasoning (or computation in general) has been one of the most prominent themes in AI, especially in terms of progress in cognitive robotics, and a considerable number of approaches have been reported. However, they are quite objective, that is, not based on a certain plausible model of human cognition of 4D world such that it can produce and understand intuitive or subjective 4D expressions in natural language. This paper describes 4D language understanding based on Mental Image Directed Semantic Theory, focusing on intuitive 4D representation and reasoning.

  • 河合 継, 島田 径汰, 長野 大知, 下辻 貴宏
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 10-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    本論文では, 対話型AIが表情を変化させることによって, コミュニケーション相手にどのような印象を与えるかを調査する. また, AIに人間と同様な表情変化をさせる新たな感情表現手法を提案する. 歌手が感情を込めて歌を歌うように, 人間は言葉を発する際に, 表情や声で感情を表現することができる. 今回は, 話し手に全く表情変化がない場合とGoogle APIを利用して取得した感情情報を用いて話し手の表情を変化させた場合, 提案手法であるイメージ情報によって話し手の表情を変化させた場合の3つの場合において, 聴き手が話し手からどのような印象を受けるか比較を行った. 比較の結果, Google APIのみを用いた手法と比べて提案手法の方が, 感情表現が豊かになることや不自然な表情を減らす効果があることが示された.

  • 金井 良太, 藤澤 逸平, 玉井 信也, 眞方 篤史, 安本 雅啓
    原稿種別: 研究会資料
    2019 年 2019 巻 AGI-013 号 p. 11-
    発行日: 2019/11/22
    公開日: 2021/09/16
    研究報告書・技術報告書 フリー

    In this paper, we propose a hypothesis that consciousness has evolved to serve as a platform for general intelligence. This idea stems from considerations of potential biological functions of consciousness. Here we define general intelligence as the ability to apply knowledge and models acquired from past experiences to generate solutions to novel problems. Based on this definition, we propose three possible ways to establish general intelligence under existing methodologies for constructing AI systems, namely solution by simulation, solution by combination and solution by generation. Then, we relate those solutions to putative functions of consciousness put forward, respectively, by the information generation theory, the global workspace theory, and a form of higher order theory where qualia are regarded as meta-representations. Based on these insights, We propose that consciousness integrates a group of specialized generative/forward models and forms a complex in which combinations of those models are flexibly formed and that qualia are meta-representations of first-order mappings which endow an agent with the ability to choose which maps to use to solve novel problems. These functions can be implemented as an "artificial consciousness". Such systems can generate policies based on a small number of trial and error for solving novel problems. Finally, we propose possible directions for future research into artificial consciousness and artificial general intelligence.

feedback
Top