成果報告書詳細
管理番号20160000000614
タイトル*平成27年度中間年報 次世代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2016/7/22
報告書年度2015 - 2015
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Achievement summary
1. Development of artificial vision system
Basic design of bi-directional deep network: Existing studies have revealed that
bottom-up hierarchical networks can acquire data-driven features that are similar
to those found in the secondary visual system of mammals, but cannot those in
further higher visual systems. To realize wide variety of visual perceptions, we
sought basic architecture of bi-directional (i.e., combination of bottom-up and
top-down) deep neural networks and their learning scheme. Based on literature
reviews, we chose ‘predictive coding’ as the basic learning scheme. Related to
this architecture design, we prepared several image datasets to be used for training,
and a cluster-based computer system to perform efficient learning.
Decoding of general objects: We presented a basic scheme for decoding arbitrary
objects from brain activities, using the machine vision principle that an object
category is represented by a set of features rendered invariant through hierarchical
processing. We showed that visual features including what a deep convolutional neural
network (CNN) can produce from human fMRI activities. Predicted features are used to
identify the target object from a set of computed features for a large number of objects.
2. Development of artificial motor control system
Nonlinear schemes of reinforcement learning and inverse reinforcement learning: We
developed a reinforcement learning (RL) scheme whose value function was represented
by a non-linear neural network. By employing a non-monotonically increasing function
as an activation function, we realized a more compact architecture than in the existing
studies, but with comparable performance, when evaluated by computer games like SZ Tetris.
We also designed an inverse reinforcement learning (IRL) scheme that combines the estimation
of likelihood ratio and the deep learning; we expected the deep learning to extract
appropriate features even from high-dimensional experience data. When our scheme was applied
to an apprenticeship learning problem of the Reversi game, in which the opponent was set as
a winner among three heuristic players, we found our apprentice agent got stronger than all
of the three opponent heuristics.
Multiple-paired architecture of reinforcement and inverse reinforcement leaning: Toward
development of multiple-paired architecture of RL-IRL modules, we first developed an efficient
prediction-based control scheme that can be applied to real-time control of non-linear,
high-dimensional systems. In particular, we developed a prediction-based scheme including a
hierarchical optimization within an eigen-perturbation space of the system. When applied to a
three-dimensional (3D) bipedal robot simulator, our new method realized not only computationally
efficient controls but also variety of 3D motions, such as running, jumping and squatting.
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。

▲トップに戻る