| SASHI NOVITASARI | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 作村 諭一(BS), Sakriani Sakti |
|
Title: Neural Incremental Speech Recognition Towards Simultaneous Speech Translation
Abstract: Real-time machine speech translation systems translate the incoming speech into the target language in real-time. This system can be achieved by performing low-latency processing in ASR before passing the output to MT and then TTS modules. However, the current state-of-the-art neural ASR with attention mechanism requires high delay because it has to analyze the whole input for an output. On the other hand, existing attention-based incremental ASR (ISR) is more complicated to construct than the standard ASR. In this work, we propose attention-transfer ISR, an ISR that learns the knowledge from attention-based non-incremental ASR for segment-based recognition. Toward a real-time machine speech translation, we evaluate the effect of AT-ISR performance on the translation task. Language of the presentation: English | |||
| SUMAILA NIGO | M, 2回目発表 | 知能コミュニケーション（ソーシャル・コンピューティング） | 中村 哲☆, 松本 裕治, 荒牧 英治, 若宮 翔子 |
|
Title:Web-based Epidemic and Pandemic-Prone Diseases Surveillance using Google Search Volume Abstract:People leave traces about their wellbeing on the Internet, and these traces can be captured and used to derive actionable information, one of such is web-based disease surveillance. Existing methods toward web-based disease surveillance fail in certain contexts, e.g., in places with highly biased data, such as in developing countries. We propose a new approach based on Google Search Volume data that can produce reliable results in multiple contexts. Language of the presentation: English | |||
| 関 直哉 | M, 1回目発表 | 知能コミュニケーション（データ駆動知識処理） | 中村 哲, 松本 裕治, 鳥澤 健太郎, 飯田 龍 |
|
title: Answer Selection in Factoid Question Answering using Sentiment Classification
abstract: Factoid question answering is a task of answering what-, who-, when-, and where-type questions. For example, “What should I present to my children in Christmas?” is given as a question and its answer is “Nintendo Switch”. Typical question answering systems first extract as answer sentence candidates, sentences or texts in which an answer might be included, and then extract noun phrases corresponding to answers. Such answer sentence candidates, however, may be inappropriate in terms of sentiment polarity (e.g., a question asks a positive aspect of something, like the above example question, but the answer sentence candidates consist of only negative contents) and the inappropriate sentence candidates lead to wrong answers. The aim of this work is to develop a factoid question answering method that can distinguish such positive/negative aspects of questions and answer sentence candidates and that can give proper answers to the questions with such sentimental aspects. For example, if a question “What should diabetics eat?”, which we assume that has a sort of positive aspect, is given, an answer “onion” in an answer sentence candidate “For diabetics, eating [onion] is effective to lower blood sugar level.” is appropriate because the answer sentence candidate has a positive polarity indicated by “is effective”. On the other hand, an answer “cake” in an answer sentence candidate “Too much sugar rich food such as [cake] may cause diabetes.” is inappropriate because this answer sentence candidate is negative. In our method, we apply a sentiment classifier to a given factoid question and its answer sentence candidate. And then we provide an answer taken from the answer sentence candidates that have the same sentiment polarity as that of the question. Also, we try to deal with the cases where the answer sentence candidates do not indicate particular sentiment polarities by looking at the surrounding text of the candidates or by looking at the answer sentence candidates that have the same answer. Note: [ ] denotes an answer is extracted from this part of an answer sentence candidate. language of the presentation: Japanese | |||
| 川崎 明宙 | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸, 末次 志郎(BS), 大竹 義人, スーフィー マーゼン, 日朝 祐太 |
|
title: Evaluation of the deformable registration using Convolutional Neural Network
abstract: Deformable registration is a method commonly used in medical research and clinical routine to analyze the pre- and post-operative images or evaluate inter-patient pathological variation. However, existing techniques using a numerical optimization at test phase requires significant amount of computation time. In order to address this problem, CNN-based methods have been proposed, which are classified into two categories; supervised and unsupervised. In this study, we employed an unsupervised method called Voxelmorph. It computes the registration field between the moving image and fixed image without using the ground truth deformation fields for the training, and notably, it also provides an estimation of the uncertainty of the deformation field at each voxel. In this research, we aim to evaluate the Voxelmorph registration using the CT image databases containing bones and teeth. The results of the registration with the conventional method (ANTs, Advanced Normalization Tools) were compared to Voxelmorph. We also discuss about a possibility of the analysis of the large-scale CT image databases and its variety of applications including multi-atlas segmentation, landmark detection, and statistical shape and density modeling. language of the presentation: Japanese 発表題目: CNNを用いた非剛体レジストレーションの大規模データベースによる検証 発表概要: 医用画像の非剛体位置合わせは同一患者の術前術後解析や，患者間の病態の比較をおこなうために医学研究や臨床現場で多く用いられている技術である．しかし，既存の実行時に最適化計算を必要とする手法は，長い計算時間を要する．これを解決するため， 近年CNNによるレジストレーション手法が注目されている．大きく，教師あり学習と教師なし学習の二つのアプローチがこれまで提案されており，本研究では教師なし学習によりCNNの学習をおこなうVoxelmorphという手法を用いた．本研究では，骨や歯などの硬組織を含む実CT画像でのVoxelmorphを用いた非剛体位置合わせの精度および大規模データベース解析における実用性の検証をおこなうことを目的とする．本発表では，Voxelmorphと，実行時に最適化計算を行う従来法（ANTs, Advanced Normalization Tools）との非剛体位置合わせの比較検証をおこなったので報告する．また，本手法を用いた大規模な頭部および骨盤データベースでの解析結果について述べ，マルチアトラスセグメンテーション，ランドマーク抽出，骨形状・骨密度の統計的解析などへの応用の可能性について議論する． | |||
| LIANG TIANHENG | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸, 末次 志郎(BS), 大竹 義人, スーフィー マーゼン, 日朝 祐太 |
|
Title: Classification of Dilated/Hypertrophic Cardiomyopathy Using Pathology Image
Abstract: Recently, the classification task on pathology images widely increase its uses. These tasks aim to support pathology doctors to review the images from patient which could save a lot of time and increase accuracy of diagnosis of diseases. This project is a 3 classes classification on Dilated Cadiomyopathy(DCM), Hypertrophic Cardiomyopathy(HCM) and Normal cell. Currently working on the task of classifing pathology images from 3 classes, using and comparing between both supervised and unsupervised learning method. language of the presentation: English | |||
| 横田 京祐 | M, 1回目発表 | ロボティクス | 小笠原 司, 佐藤 嘉伸, 高松 淳, 丁 明 |
|
Title:Development of A System that detect Adnormalities during Walking and Prevents Falls
Abstract:In Japan, where the birthrate is declining and the population is aging, how to deal with falls, which is one of the causes of bedridden, is an issue. Therefore, will develop a system for posture measurement using ZMP, warning by voice, and posture assistance by walking stick. The ZMP measurement uses a force sensor on the sole of the foot, and assist with a stick is performed by motion control using wheels. language of the presentation: Japanese | |||
| 小坂 麻人 | M, 1回目発表 | 知能システム制御 | 杉本 謙二, 安本 慶一, 小林 泰介 |
|
title: Emergence of Partnership by Mutual Understanding of Values of
Human and Robot
abstract: Nowadays, robots work only for humans who give the tasks to be conducted explicitly. Their tasks, however, tend to have no cooperation with humans for security and simplicity. Towards next-generation robots in human-robot symbiotic society, they are required to gain capabilities to physically interact with humans. This study, therefore, develops a new multi-agent system with humans, where all the agents are supposed to be controlled by reinforcement learning with their own objectives (i.e. rewards). To emerge cooperative behaviors, the mutual communication of rewards is the most important technology. This presentation considers the way to do so and shows the performance of such a multi-agent system. language of the presentation: Japanese | |||
| 西 陽太 | M, 1回目発表 | ソフトウェア工学 | 松本 健一, 中島 康彦, 石尾 隆, Raula G. Kula |
|
title: Equivalence verification of ported program using symbolic execution
abstract: Hardware support is important to improve the performance of computationally intensive algorithms such as deep learning. In particular, CGRA is getting attention as a power-saving and high-performance hardware architecture. To utilize its hardware features, programmers have to manually port a regular C program to a CGRA-enabled C program with special instructions provided by the hardware. It is necessary to verify the equivalence of the programs because the rewriting process is error prone. This study proposes a method to verify the equivalence of such a ported program and the original version using symbolic execution. Symbolic execution is a program analysis technique which treats input values as a symbol instead of concrete values, and extracts executable paths and obtains constraints to each path through symbol operations. We use this to check that the ported function returns the same result as the original version. language of the presentation: Japanese | |||
| 馬越 圭介 | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, 岡田 実, 藤本 まなと, 松田 裕貴 |
|
title: Investigation on Indoor Positioning Estimation Method Using Doppler Sensor.
abstract: Recently, device-free recognition proposed that can infer human activity without invading resident privacy and attaching sensor. In this study, we focus on a device-free human activity recognition system using doppler sensors and propose and examine a new position estimation method using doppler sensors. Until now, position estimation using doppler sensor is difficult to accurately estimate the position of the object due to various factors such as "distance between sensor and object" and "difference in stationary/moving state of the object". Hence, in this study, we examined whether the position of the object can be accurately estimated by mathematical modeling the positional relation between the output signal of the Doppler sensor and the object. As a result, we would conclude that the distance between the object and the sensor can be estimated if they are apart from each other within 200 cm. language of the presentation: Japanese 発表題目:ドップラーセンサを用いた位置推定手法の検討 発表概要: 近年，居住者のプライバシーを侵害することなく，居住者が宅内で何も身に付けずに，行動を推定可能であるデバイスリーによる行動認識システムが提案されている．本研究では，ドップラーセンサによるデバイスフリー行動認識システムに着目し，ドップラーセンサを用いた新たな位置推定手法の提案・検討を行う．これまで，ドップラーセンサを用いた位置推定では，「センサと対象物との距離」や「対象物の静止/運動状態の違い」などの様々な要因により，対象物の位置を正確に推定することは困難であった，そこで，本研究では，ドップラーセンサの出力信号と対象物の位置関係を数理モデル化することによって，対象物の位置を正確に推定可能かどうかの検討を行なった．その結果，対象物体とセンサの距離が200cm程度であれば距離が推定可能であると示唆された． | |||
| 片山 洋平 | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, 向川 康博, 伍 洋（客員）, 諏訪 博彦, 藤本 まなと |
|
title: Consideration of sightseeing movie curation system using drive recorder movie
abstract: In recent years, the number of users who use video is increasing when planning tourism because of the spread of SNS.However,A tourist can find sightseeing spots video in Internet, but there are not many that include the route to the sightseeing spot. In this Research, we focus on drive recorders that are widely used for crime prevention, and consider algorithms for curating (summarizing) video obtained from them as sightseeing videos. When curation was performed using speed information, location information of sightseeing spots, and color histograms between frames, which are part of the algorithm of the proposed method, the playback time could be reduced to 6.8% of the original video. language of the presentation: Japanese 発表題目:ドライブレコーダ動画を用いた観光動画キュレーションシステムの検討 近年，SNSの普及により観光の計画を立てる際に動画像を用いるユーザが増加している．しかし，アップロードされる動画には観光地の映像は多数あるが，観光地までの経路が含まれるものは多くない．そこで本研究では防犯用として普及が広がっているドライブレコーダに着目し，そこから取得される動画像を観光動画としてキュレーション(要約)するアルゴリズムの検討を行う．実際に提案手法のアルゴリズムの一部として用いられる，速度情報と観光地の位置情報，フレーム間のカラーヒストグラムを利用しキュレーションを行なったところ元動画の6.8%の再生時間に短縮することができた． | |||
| CONG XI | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, 向川 康博, 酒田 信親, 磯山 直也 |
|
title: A fire simulation system to raise risk awareness of fire using Augmented Reality and Material identification
abstract: Many people have not noticed the risk of fire. The reason is that they do not have an impressive fire experience. In this study, we will use Augmented Reality to let users experience fire in the space they lives everyday. we will also use material recognition to let the flame show different shapes, smoke, and diffusion depending on the object`s material in the room. language of the presentation:Japanese | |||