| 木村 睦 | 1561010: D, 中間発表 | コンピューティング・アーキテクチャ | 中島 康彦,池田 和司,高前田 伸也,TRAN THI HONG |
|
title: Brain-type Integrated System using Thin-Film Devices
abstract: Artificial intelligences are promising as exceedingly important technologies in future information societies. However, current ones are realized using complicated programs executed on server computers, whose size and power consumption are unbelievable large. Therefore, we are developing "Brain-type Integrated System using Thin-Film Devices", which mimic the computing architecture of human brains in the hardware level. Because thin-film devices can be fabricated in the three dimensional stacked structure using low cost processes including printing techniques, they seem to be key devices for the brain-type integrated systems. It is expected that the artificial intelligences can be compact and low power similar to living brains and put on everything, which leads to Internet Of Things. Currently, we are trying to make a sample device and confirm basic operations: First, we succeeded in simplifying the processing elements, such as neuron and synapses, which can increase the number of the processing elements integrated in the system. Next, we proposed cellular neural networks where a neuron is connected to only the neighboring neurons, which are suitable for composing using electronic circuits. Moreover, we confirmed that some kinds of thin film, such as amorphous In-Ga-Zn-O and Ga-Sn-O, have a preferable characteristics for synapses employed for unsupervised learning. Now, we are trying to confirm the correct working of letter reproduction using a cellular neural composed of the simplified processing elements and thin-film devices. At the end of the doctoral course, I would like to develop a brain-type integrated system having three dimensional thin-film elements and confirm more complicated functions including image recognition, etc. language of the presentation: Japanese | |||
| 岡田 裕斗 | 1551025: M, 2回目発表 | 数理情報学 | 池田 和司,中村 哲,川人 光男,森本 淳 |
|
title: The estimation of deep brain activity using resting state fMRI and fNIRS
abstract: Functional magnetic resonance imaging (fMRI) and functional near infrared spectroscopy (fNIRS) are methods for non-invasively measuring human brain activity. However, fMRI restricts subject's behavior, and fNIRS can't measure deep brain activity. Therefore, we develop better measurement methods by combining multiple measurement methods. In this study, I develop the method to estimate deep brain activity from brain surface activity acquired by fNIRS. It's known to be extracted multiple brain networks by resting state fMRI (rs-fMRI). Hence, I verify whether to generalize between tasks and between subjects by using information of the brain networks in the resting state. As the result, generalization of the estimation between tasks and between subjects is possible to some extent. In addition, I associate channels of fNIRS with brain region of fMRI in order to combine fNIRS and fMRI. language of the presentation: Japanese | |||
| LAO BRYAN TSANG | 1551127: M, 2回目発表 | 数理情報学 | 池田 和司,中村 哲,爲井 智也 |
|
title: Towards the Development of Automated Sit-to-Stand Rehabilitation abstract: The sit-to-stand (STS) movement is a functional task that is a pre-requisite to most activities of daily living. Thus, its difficulty affects the quality of life of many elderly people. Understanding effective STS movement is essential for improving rehabilitation strategies and developing services for the rapidly increasing number of elderly people. This study aims at identifying effective STS therapy by analyzing the movements induced by therapists of different degrees of skill. Mathematical models were constructed and their parameters were identified with experimental data and physiological phenomena. We aim to use these results to understand effective STS therapy and to develop an automated rehabilitation system. language of the presentation: English | |||
| 古庄 泰隆 | 1551096: M, 2回目発表 | 数理情報学 | 池田 和司,松本 裕治,久保 孝富 |
|
title: Roles of Pre-training in Deep Neural Networks from Information Theoretical Perspective
abstract: Pre-training is used for parameter initialization method of deep neural networks (DNN) before training. Although it is known that pre-training improves performance of the DNN, relationship between properties of the DNN before and after training and relationship between the property after training and the performance are unclear. In order to clear these relationships, we considered DNN as encoder and evaluated information theoretical measures of representations of last hidden layer as the property of the DNN. We found that the information theoretical measures before and after training were positive correlated and the information theoretical measures after training and the performance had a relationship. language of the presentation: Japanese | |||
| 田口 智之 | 1551057: M, 2回目発表 | ロボティクス | 小笠原 司,横矢 直和,高松 淳,丁 明 |
|
title: Boundary detection in stereo vision abstract: Mobile robots need to recognize the environment for safe and smooth navigation. Environment recognition is to organize environmental structures such as walls, floor, and obstacles, as well as understanding the relation between them. In this study, we detect the boundary lines using stereo vision. This is the first step for organizing the structure. We propose to use both forward dynamic programming and backward dynamic programming for stereo matching and to extract the boundary descriptor from the cost calculated by the dynamic programming. Then we distinguish boundaries using linear Support Vector Machine. We show the tentative result of the proposed method. language of the presentation: Japanese | |||
| 福井 友季也 | 1551087: M, 2回目発表 | ロボティクス | 小笠原 司,横矢 直和,高松 淳,丁 明 |
|
title:
Semi-Autonomous Unmanned Aerial Vehicle based on First Person View Motion Commands
abstract: Recently, unmanned aerial vehicles (UAVs) are attracting attention in the fields of infrastructure inspection and security. Controlling UAVs in complex environments using the limited information of FPV is a challenge for operators. This is because they need to avoid collisions with obstacles and verify the UAV state simultaneously. In this research, a semi-autonomous system to operate an UAV by only giving motion commands relative to the FPV is proposed. The proposed system consists of an autonomous flight component with collision avoidance, and an intuitive, easy to use user interface. With this system, operators will not have to intensively maneuver the UAV nor to consider the complexity of the environment, which will allow them to focus on the task itself. language of the presentation: Japanese | |||
| 白川 誠 | 1551052: M, 2回目発表 | 視覚情報メディア | 横矢 直和,小笠原 司,佐藤 智和,河合 紀彦 |
|
title: Free-viewpoint omnidirectional cinemagraph generation
abstract: Recently, applications such as Google Street View that enables us to explore the virtual space created based on real images are becoming common. Although such applications currently display still images of a target scene, it is expected to give users more highly realistic sensation using videos that can show motions of moving objects. However, to achieve such applications, there are two problems. As the first problem, the amount of data stored in a database becomes enormous. If videos are short, users feel discontinuous between the last and first of the looping videos. To solve this problem, we employ a cinemagraph, which is an infinite looping video in which some objects continuously move. As the second problem, recording every scene takes lots of time. For reducing the cost, we generate a cinemagraph at a viewpoint by using free-viewpoint image rendering. Therefore, our work develops a method for omnidirectional cinemagraph generation at any point of view. language of the presentation: Japanese 発表題目: 任意視点における全方位シネマグラフ生成 発表概要: 近年，Google Street View のような現実環境の情報を用いたテレプレゼンスシステムが普及している。現状，このようなシステムでは静止画像を提示しているが，より高い臨場感を得るためには，動物体の動きを再現できる動画像を用いるほうが望ましい。しかし，動画像を用いたシステムを実現するためには問題点が二つある。一つ目に，そのまま動画像を用いると，膨大なデータ量となる点である。また，データ量を減らすために動画像の尺を短くすれば，長時間再生時にループによる不連続が生じる。この問題を解決するために，シネマグラフと呼ばれる, 動物体のみを不連続なく無限にループさせる動画を用いる。二つ目に，撮影コストが高い点である。散策できるようにすべての地点で動画撮影を行うと，膨大な撮影量となる。この問題に対し，自由視点画像生成手法を用いて他の地点の見えを再現することで撮影量を削減する。よって，本研究では任意の視点における見えを再現したシネマグラフを生成する手法の開発に取り組む。 | |||
| 今西 雅美 | 1551014: M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸,小笠原 司,大竹 義人,横田 太 |
|
title: Improvement of hierarchical multi Atlas method for the muscle area automatic extraction from the lower extremities the entire CT images
abstract:Automated muscle Segmentation from 3D CT data of lower extremities with varying joint postures using hierarchical multi-atlas method has been proposed in my laboratory. When the patients lies on their backs, their postures are difference because their bones is deformated or they have pain.But previous method hasn't considered this. At first, I tied to divide right limbs from left ones. To do that, we can save much time and get good segmentation results. I’ll try to divide each bones to get better results. To do that, we can apply all lower extremities with varying joint posture. language of the presentation: Japanese | |||