成果報告書詳細
管理番号20160000000613
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)マルチモーダルコミュニケーション 多様な時系列情報に対する深層学習基盤の開発
公開日2016/7/27
報告書年度2015 - 2015
委託先名株式会社Preferred Infrastructure
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:Next Generation Robot Core Technology Research Project/ (Next Generation Artificial Intelligence) Multi-modal Communication/ The development of the Deep Learning framework for a variety of time-series multi-modal information (FY2015-FY2016) FY2015 Annual Report
We build a prototype robot having a function of the multi-modal sensing, and apply the deep learning technology to them. We focus on the multi-modal communication especially by language. The framework will share the information via network, store data, learn, and delive the learning model to them to achieve autonomous communication by robot. In this research, we defined 4 elements to achieve communication as follows, "Perception", "Accumulation", "Learning", "Feedback". We worked for survey and research to carry out screening and verification of the elemental technologies for multi-modal communication.
1. Perception
We made a prototype device with a multi-modal interface for the future field experiments. The prototype device had a capacity to have various inputs like sensory organs and 1TFLOPS computing performance to execute deep learning processes on it. We designed it based on the distributed computing architecture called “edge-heavy computing”, all of the data is able to process information on the device instead of sending whole data to servers.
2. Accumulation
We researched the latest technologies to accumulate, share and manage data focused on cloud-based ones and protocols to communicate among each devices. We also assumed the typical inputs of multi-modal communication such as video, audio and sensors. And prototyped a cloud-side application to manage the data.
3. Learning
We researched the existing technologies especially related to the multi-modal learning based on Deep Learning. We also implemented the latest method of video and audio recognition, and conversational model. It was proved that "chainer" implementation was effective to reproduce the paper results.
4. Feedback
We researched the distributed Deep Learning method to share the learning results among devices. As a result, it was found that the parameter server is one of trends on the distributed learning.
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