コロキアムB発表

日時: 9月24日(火)3限(13:30~15:00)


会場: L1

司会: 樫原茂
黒木 琢磨 M, 2回目発表 数理情報学 池田 和司, 松本 裕治, 浦岡 行治 (MS), 吉本 潤一郎
title: Examination of Driver's State Estimation Methods with Driver's Datasets Collected from Smartphone
abstruct: In this study, we examine a driver state estimation method that detects the driver's driving state (whether there is drowsiness or aggression). Some studies use the driver's facial expressions and movements or pulse, we use open datasets collected from smartphone apps for the feasibility study. There are three methods to consider: Linear discriminant analysis, Supervised DL, and Unsupervised DL (TCL) + linear discriminant analysis, which is basic supervised multi-class classification applications. Each method will be evaluated in terms of accuracy, generalization performance, or F-number.
Japanese
発表題目: スマートフォンで得られる程度の車載データを用いたドライバ状態推定手法の検討
発表概要: 本研究ではドライバーの運転状態(眠気や攻撃性などの有無)を検知するドライバ状態推定の手法の検討を行う。データとしてドライバーの表情や動作、あるいは脈拍を利用するものなどがあるが、本研究ではスマートフォンアプリから収集されたオープンデータセットを用い、feasibility studyを目的とする。検討する手法は、基本的な教師あり多クラス分類の応用である、線形判別分析、Supervised DL、そしてUnsupervised DL(TCL)+線形判別分析の3つである。それぞれの手法を精度や汎化性能、あるいはF値といった点から評価していく。
 
二又 航介 M, 2回目発表 知能コミュニケーション 中村 哲, 松本 裕治, 須藤 克仁
title: Data augmentation for simultaneous interpretation corpora from bilingual corpora using the technique of style transfer.
abstract: Simultaneous interpretation is a task to translate partial translation results of the target language before the input text is completed. In communication via the simultaneous interpretation system, the delay in translation becomes a major obstacle for smooth communication, so it is necessary to translate partially and accurately while minimizing the delay. In particular, simultaneous interpretation between languages such as English and Japanese, which have a large difference in word order, has a major problem of delay before translation starts. On the other hand, if a translation can be performed in a form close to the word order of the source language, the delay can be reduced. bilingual corpora are usually used to train a simultaneous interpretation system as it is like used to train a machine translation system. However, the simultaneous interpretation corpus is different from the bilingual corpus used for the machine translation system. It is a parallel corpus composed of sentences in which the target language is partially translated before the input sentence is completed. Therefore, if the simultaneous interpretation corpora can be used to train a simultaneous interpretation system, the input sentence can be divided into small parts and sequentially translated, so the delay for translation would be reduced. However, since the number of simultaneous interpretation corpora currently available is few. so in this paper, we propose a method to augment simultaneous translation corpora from bilingual corpora. The proposed method uses style transfer technique to transfer a style of machine translation to one of simultaneous interpretation. we also examine the current problems arises in pseudo-simultaneous interpretation corpora generated by style transfer.
language of the presentation: Japanese
発表題目: スタイル変換技術による対訳コーパスから同時通訳への拡張
発表概要: 同時通訳とは,入力文章が完結する前に目的言語の部分的な翻訳結果を訳出するタスクである.同時通訳システムを介したコミュニケーションでは,翻訳の遅延が円滑なコミュニケーションの大きな障害となるため,遅延を最小限にしつつ正確に部分訳出をする必要がある.特に英語と日本語のように語順が大きく異なる言語間の同時通訳では,訳出開始までの遅延が大きな問題となる.一方で,原言語の語順に近い形で訳出を行うことができれば,遅延を少なくすることができる.同時通訳システムの学習には通常,機械翻訳システムと同様に対訳コーパスが用いられる.同時通訳コーパスは,機械翻訳システムの学習に用いられる対訳コーパスと異なり,入力文が完結する前に目的言語の部分訳出を行った文から構成される対訳コーパスである.したがって,同時通訳システムの学習に用いられる対訳コーパスとして,同時通訳コーパスを用いることができれば,入力文を小さな部位に区切り逐次訳出できるため,訳出を終えるまでの遅延が少なくなる.しかし,現在利用可能な同時通訳コーパスの量は非常に少ないため,このような問題設定は現実的ではない.そこで本稿では,機械翻訳に用いられる対訳コーパスから,同時通訳コーパスへと拡張する手法について提案する.提案手法ではスタイル変換を用いることで,機械翻訳のスタイルから同時通訳へのスタイルへと変換を行う.また,スタイル変換により生成された疑似同時通訳文について現状での問題点について検討する.
 
柴田 大作 D, 中間発表 知能コミュニケーション 中村 哲☆, 松本 裕治, 荒牧 英治, 田中 宏季
title: Factuality Analysis of Pain Expressions from Clinical Notes
[Background] Sharing information for pain between medical staffs and pain management of patients are significant in medical domain. However, information for patient’s pain is generally written as a free text in clinical notes. Therefore, it is extremely challenging to extract it from clinical notes. In this study, we attempt to extract pain expressions from clinical notes and conduct factuality analysis of them with Natural Language Processing technique. [Material] Clinical notes are obtained from The University of Tokyo Hospital which were generated in July, 2016. These are divided into sentences based on space, indention and comma. and 10, 000 sentences are randomly selected from them. Medical experts annotate pain expressions in the data. [Methods] We apply machine learning technique to extract pain expressions from the data. Our machine learning model use a Bi-LSTM CRF model that is composed of embedding layer, Bi-directional Long Short-Term Memory and Conditional Random Fields. [Result] Our model achieved the maximal average F1 score of 52.48 (Maximum value is 87.12). [Discussion] The experimental results show that our model could extract pain expressions with the practical accuracy.
language of the presentation: Japanese
 

会場: L2

司会: Tran Thi Hong
石坂 守 M, 2回目発表 ディペンダブルシステム学 井上 美智子, 太田 淳(MS), 中島 康彦, 大下 福仁
title: Fault-Tolerant Design for Memristor-Based Neural Network
abstract: As the matrix-vector product is the most essential operation in the weight calculation of deep learning, it has a great impact on calculation speed and power consumption of neural network circuits. A memristor is one of the most promising materials to perform an efficient matrix-vector product. However, it has been pointed out that the memristor has severely low write endurance limitation and large manufacturing variability due to its manufacturing immaturity. Accordingly, we propose a lifetime enhancement method for the memristor-based neural network (MNN). Our method consists of two-step: quantization and error correction code (ECC). The quantization contributes to reducing the impact on the process variability, and the ECC conventionally used in the conventional digital circuit can work to correct errors. Through numerical experiments, we experimentally confirmed the quantization successfully correct error under the large manufacturing variation.
language of the presentation: Japanese
 
倉角 哲也 M, 2回目発表 ロボティクス 小笠原 司, 太田 淳(MS), 加藤 博一, 高松 淳, 丁 明
title: Deep-learning-based simultaneous estimation of joint angles and torques based on the upper-arm deformation using a distance sensor array
abstract: Measure the motion of upper limbs can make more intuitive the operation for robots devices and facilitate the motion analysis of daily-life activities. Today, various biosignals have been measured to estimate the human motions. In this research, we focus on the measurement of the upper arm deformation which can provide complex motion information from different tissues including the muscles, the tendons, and the bones. In this presentation, a novel method is proposed to estimate the elbow joint angles and torques simultaneously only by measuring the upper arm deformation with a distance sensor array. Based on the measured information, the joint angles and torques can be estimated in some deep-learning-based method, such as SVR, SVM, DNN, and so on. In the experiments, we estimate the motion and evaluate the performance of the proposed method.
language of the presentation: Japanese
 
由井 朋子 M, 2回目発表 ロボティクス 小笠原 司, 佐藤 嘉伸, 高松 淳, Gustavo Garcia
title: Measurement of Dental Hygienist Hand Scaling Motion and Quantitative Evaluation
abstract: A dental hygienist is an occupation that maintains oral health and provides medical assistance. In Japan, in order to get this job, a student of the dental hygienist should learn in a training school for more than 3 years. In this school, one of the skills that students have to learn is to operate the instruments precisely in the oral cavity. However, such precise operation is not easy to be acquired without effective repetitive training. The purpose of this study is to measure and evaluate the behavior of dental hygienists quantitatively. It can develop the standardization of teaching points, and self-study tools to enable efficient and repetitive training in the future. In this research, we focus on the hand scaling operation, which is one of the most important works of dental hygienists. A measurement system is developed to measure the operating force and motion simultaneously using a force sensor and motion sensors. The skill could be evaluated by comparing the measured data between beginner and expert. In this presentation, I will introduce the developed system and show some the measured results in the pre-experiment for extracting the characteristics of the scaling motion.
language of the presentation: Japanese
 
中野 萌士 M, 2回目発表 サイバネティクス・リアリティ工学 清川 清, 加藤 博一, 酒田 信親
title: Impact on Vision-induced Gustatory Manipulation
abstract: This presentation proposes a novel gustatory manipulation interface which utilizes the cross-modal effect of vision on taste elicited with AR-based real-time food appearance modulation using a generative adversarial network (GAN). Unlike existing systems which only change color or texture pattern of a particular type of food in an inflexible manner, our system changes the appearance of food into multiple types of food in real-time flexibly, dynamically and interactively in accordance with the deformation of the food that the user is actually eating by using GAN-based image-to-image translation.
language of the presentation: Japanese