成果報告書詳細 |
---|

管理番号 | 20170000000691 |

タイトル | *平成28年度中間年報 次世代人工知能・ロボット中核技術開発 次世代人工知能技術分野 計算神経科学に基づく脳データ駆動型人工知能の研究開発 |

公開日 | 2017/8/30 |

報告書年度 | 2016 - 2016 |

委託先名 | 株式会社国際電気通信基礎技術研究所 |

プロジェクト番号 | P15009 |

部署名 | ロボット・AI部 |

和文要約 | |

英文要約 | Title: Strategic Advancement of Multi-Purpose Ultra-Human Robot and Artificial Intelligence Technologies (SAMURAI)/Next-generation AI technology/Development of brain data-driven artificial intelligence based on computational neuroscience (FY2015-FY2017)
Achievement summary 1. Development of artificial vision system Bidirectional deep network: To realize wide variety of visual perceptions, we have been seeking the architecture of bi-directional (i.e., combination of bottom-up and top-down) deep neural networks and their learning scheme. We have designed a basis learning algorithm for nonlinear bidirectional deep networks based on the theory of mixture of sparse coding models. We achieved an efficient learning procedure based on independent component analysis. Examining the receptive fields of our bidirectional deep network, we have found that our network well reproduces properties of face-selective neurons in the primate’s inferior-temporal cortex. Decoding of general objects: Using the scheme we proposed last financial year, we have demonstrated that visual features derived from a deep convolutional neural network can be predicted from fMRI activity patterns, and that greater accuracy is achieved for low/high-level features with lower/higher-level visual areas, respectively. Predicted features are used to identify object categories (extending beyond decoder training) from a set of computed features for a large number of object images. Our results demonstrate a homology between human and machine vision and its future utility for brain-based information retrieval methods. 2. Development of artificial motor control system Nonlinear schemes of reinforcement learning and inverse reinforcement learning: To handle huge state space tasks by reinforcement learning, we have developed novel non-monotonically increasing activation units. One is a Sigmoid weighted Linear (SiL) unit that is a product of the sigmoid and linear functions and the other is its derivative (dSiL) unit. They are used with a standard convolution neural network to represent a state-action value function. Experimental results on Atari 2600 games has shown that our proposed network achieves better performance than the original Deep Q Net (DQN) and its later extensions. Next, we formulate inverse reinforcement learning as density ratio estimation between state transition probabilities before and after learning. This is solved by binary logistic regression in which the classifier is represented by the immediate reward and state value function, so that the classifier is given by deep neural networks. As a consequence, we have found our method can extract nonlinear reward representation. Parallel and hierarchical architecture for motor control system: We have developed a parallel and hierarchical architecture and s control software platform for implementing real-time model predictive control on a humanoid system. According to this architecture, model predictive control calculation is conducted on a high-performance calculation environment while the derived control output by the model predictive control is properly executed on the real humanoid robot to generate simple sinusoidal lower-limb movements. Through simple oscillatory movement tasks, we have confirmed this new architecture and software allow the high-dimensional humanoid system to perform real-time control. |

ダウンロード | 成果報告書データベース（ユーザ登録必須）から、ダウンロードしてください。 |