ゼミナール発表

日時: 12月6日(水)3限(13:30-15:00)


会場: L1

司会: 油谷 暁
金谷 勇輝 1751030: 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
title: Development of tour video curation system toward intuitive sightseeing tour recommendation
abstract: Recently, there are increasing demands for tour recommendations using videos because more tourists search and watch the tour videos when planning a sightseeing tour. However, searching desired videos will be very hard (even not possible) or require significant labor. In addition, since each searched video is on a fixed route, it is necessary to search other route videos when it does not match the user's desire. In this study, we develop a system for curating a tour video adjusted to the intention or preference of each individual user from consumer generated media (CGM) including videos, photos and comments posted to each sightseeing spot/route through SNS.
language of the presentation: Japanese

 
佐々木 渉 1751050: 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
title: Daily Living Activity Prediction in Smart-Home based on Activity Duration Classification
abstract: Recently, a lot of daily living activity recognition methods using sensors, cameras and/or smartphones are proposed. For example, there is a method that recognizes and visualizes daily living activities of elderly people by using smartphone. However, most of the existing studies focus on recognizing the current user activity. Future activity prediction is needed when providing more sophisticated services such as making bed room temperature comfortable just before the user uses the room. In this study, we will propose a future activity prediction method that focuses on the duration of preceding activity.
language of the presentation: Japanese

 
高田 将志 1751060: 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
title: A Smartphone-based Core Muscle Training Support System
abstract: Core Muscle Training (CMT) is attracted as a practical and easy training for the person who lacks an exercise. To maximize the effect of training, a user needs to keep a correct posture. However, it is difficult to recognize own posture and the difference with the correct posture during training. In this research, we propose a smartphone-based CMT support system which measures the body posture by using built-in inertial sensors. It compares between the measured posture and ideal posture and provides feedback to lead to the correct posture. It also evaluates workout effectiveness from "body shivering" information obtained from the sensors.
language of the presentation: Japanese

 
高城 賢大 1751063: 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
title: Prediction of drowsiness and sleep support for low function autism based on life logs and environmental information
abstract: More than 50% of autistic patients have sleep disorders, which is a heavy burden for patients and their families. However, it is difficult to detect drowsiness of low-function autistisms using a heart rate monitor or an electroencephalograph because they don't like attaching sensors such as wearable bands. Therefore, in order to support low function autistic disorder with sleep disorder, we create a drowsiness prediction model from life log data and environmental information, and develop a sleeping support system.
language of the presentation: Japanese

 
AKPA AKPRO ELDER HIPPOCRA 1661022: D, 中間発表 ユビキタスコンピューティングシステム 安本 慶一, 佐藤 嘉伸, 荒川 豊
title: In or Out: Calories Intake and Expenditure Estimation Systems
abstract: Today, I will introduce two of my latest works about using IoT and ubiquitous systems to support healthy-life. As you might know, the amount of calories in human-body is very related to human wellbeing. Measuring and assessing intake and burned calories is required in the treatment of many chronic diseases. To help in the assessment of daily in and out calorie, we developped two systems: a) A food calorie estimation system by using smartphone camera and eating tools like chopsticks; b) A smart fitness glove that uses pressure sensors to monitor burned out calories during fitness activities.
language of the presentation: English
 

会場: L2

司会: Juntao Gao
中井 文哉 1751076: M, 1回目発表 数理情報学 池田和司☆
 
中原 英里 1751084: M, 1回目発表 数理情報学 池田和司
Title:Gaze Behavior of Dog Experts and Non-experts Differs using Watching Dog Training
Abstract: Humans can interpret social signals from not only other humans but also other species with sociality such as domestic dogs.However, little is known about how they detect and understand social signals from other species, which may reflect the evolution of communication and sociality. Since humans get informationfrom vision and the specialty often appears in gaze, we investigated the gaze behavior of experts and non-experts of dogs while they watched video clips in which an expert or a non-expert trained a dog and detected dogs’ social signals using an eye-tracker. I report preliminary results with some discussion.
Language of the presentation: Japanese
 
久田 将史 1751099: M, 1回目発表 数理情報学 池田和司
Title: Analysis of Driver's Gaze Behavior in Winding Road
Abstract: Driving behavior prediction is essential to develop an advanced driver assistance system(ADAS) or self-driving cars since they should understand how surrounding vehicles move into the next step. One of the ways to understand driving behavior is to measure their gaze, since drivers perceive the road ahead and control the steering wheel so that the vehicle trace the line the driver plans in mind. When a driver drives on a curve, the driver's gaze is mainly focused around a tangential point. However, especially on winding roads, driver's gaze strategy is controversial since the line is unseen. Therefore, in this research, we investigated driver's gaze in winding roads. We have collected gaze data during driving on a real road using an eye tracker. In this presentation, I will report a preliminary result.
Language of the presentation: Japanese
 
M. ROSYIDI 1751134: M, 1回目発表 数理情報学 池田和司

title: *** M-Statistic Based Change Point Detection for Intelligent Transportation System ***

abstract: *** To improve road safety with intelligent transportation system, detection of abnormal events on the road is an indispensable technology.We propose a change point detection using Maximum Mean Discrepancy (MMD) for the detection of such abnormal events. MMD measures the discrepancy between a pair of distributions without density estimation,and need few assumptions on the distribution. We investigated whether our proposed method can detect abnormal events in a dataset obtained with a transportation simulator. I will report brief preliminary results in this talk. ***

language of the presentation: *** English ***

 
藤谷 拓矢 1751103: M, 1回目発表 ディペンダブルシステム学 井上 美智子
title: title:Paper introduction Distributed Evacuation in Graphs with Multiple Exits
abstract: In case of a disaster, you have to evacuate from building, subway, etc. Each costomer there is thought to move to the closest exit, so the number of pepole is different between the exits: there are crowded and vacant exits. Forcusing on it, I'll introduce Evacuation problem described as graph. There are several exit and some argent on the graph. Evacuation problem is problem evacuating all agent using exit from the graph .
language of the presentation: Japanese
: タイトル:論文紹介 Distributed Evacuation in Graphs with Multiple Exits
: 概要:ビルや地下で災害が発生した場合,非常出口を使って避難する必要があります. その場合,最も近い非常出口から避難しようと試みますが,他の人たちも同様に動くと混み合う出口とそうでないものとができる可能性があります. 今回はそれに着目し,グラフでモデル化した避難問題を紹介します. 避難問題は,複数の出口が設置されたグラフ上で,エージェントを出口に移動させ,グラフから全てのエージェントを取り除く問題です. 論文で得られたいくつかの結論を紹介した後,紹介されていたアルゴリズムを紹介します.
 
LIU YING 1751133: M, 1回目発表 モバイルコンピューティング 伊藤 実
title: Application of Back-Pressure Algorithm to Traffic Signal Control in Road Networks of Finite Road Capacity
abstract: Back-pressure algorithm has been increasingly attractive to reduce traffic congestion for road networks. Recent work has shown the performance superiority of back-pressure based traffic signal control algorithms. However, these back-pressure based traffic signal control algorithms either assume each road can hold infinite vehicles or need to have prior knowledge of vehicle turning ratios, all of which are not realistic for applications. In this research, we propose a back-pressure based traffic signal control algorithm that can effciently reduce traffic congestion, and thus vehicle delay, for realistic road networks with finite road capacity and without prior knowledge of vehicle turning ratios. As validated by simulations, our algorithm reduces average vehicle delay by 66.7% under moderate vehicle arrival rate when compared to fixed cycle traffic signal control.
language of the presentation: Japanese
 

会場: L3

司会: 能地 宏
中村 良 1751087: M, 1回目発表 知能コミュニケーション 中村 哲
title: Rethinking Attention
abstract: Memory Network and Transformer have the possibility that high expressive power may be obtained by separating the Memory into Key and Value and by non-trivial transformation between Key-Value pair. However, many attentive networks including them use the same input matrix or linear transformation of the same input matrix for Key and Value. This lacks flexibility because the nonlinear transformation between Key-Value pair is weak and does not depend on the context other than the input itself. In order to realize a more nonlinear and context-dependent transformation for the post-branch encoding of the same input matrix, we used a Mixture of Experts using multi-layer networks as experts and a Query (ie, context) as a input of gating network.
language of the presentation: Japanese
 
中山 佐保子 1751088: M, 1回目発表 知能コミュニケーション 中村 哲
title: Code-switching ASR with DNN
abstract: Alternating the language in one conversation is called code-switching. Since the current automatic speech recognition is composed of models optimized for each language, it can not recognize the conversation with code-switching. In this research, we focus on code switching in Japanese and English, and we propose the automatic speech recognition capable of recognizing code-switching by using end-to-end deep learning based on attention model.
language of the presentation: Japanese
発表題目: 深層学習を用いたコード変換の音声認識
発表概要: 一つの会話の中で言語が切り替わるものをコードスイッチングという。現在の音声認識は、各言語に最適化されたモデルで構成されているためコード変換を含む会話を認識できない。そこで本研究では、日本語と英語のコードスイッチングに焦点を当て、アテンションモデルをベースにしたエンドツーエンドの深層学習を用いてコードスイッチングを認識できる自動音声認識を提案する。
 
西村 優汰 1751091: M, 1回目発表 知能コミュニケーション 中村 哲
title: Multilingual translation with unenven corpora
abstract: A large number of bilingual corpus is necessary for machine translation. However, depending on the language pair, the amount of data may be extremely small. There is a multilingual translation as the solution of low-resource language pairs, but the amount of data often varies depending on the language even in the multilingual corpus. Therefore, in this research, we focus on the corpus which has multiple languages but also the uneven data, and propose multilingual machine translation using uneven data.
language of the presentation: Japanese
 
古川 智雅 1751105: M, 1回目発表 知能コミュニケーション 中村 哲
title: Building a dialogue system that clarifies ambiguities in user utterance
abstract: The number of applications which communicate with human is increasing. It is necessary that the system of such applications understand natural language, but the analysis of dialogue sentence often has ambiguities. In this work, we focus on the point that the system can generate question response for utterance in spoken dialogue if they have ambiguity, and propose a dialogue system which generate question response for utterance including ambiguity.
language of the presentation: Japanese
発表題目: 曖昧性を持った発話に対して聞き返しを行う対話システム
発表概要: 近年、音声対話を行うアプリケーションが増加している。そういったアプリケーションではシステムが人間の言葉を理解することが必要だが、音声入力文の解析結果にはしばしば曖昧性が含まれる。そこで本研究では、音声対話ではユーザの発話に解析が難しい曖昧性があればユーザに聞き返すことが可能な点に着目し、曖昧性を持った発話に対して聞き返しを行う対話システムを提案する。
 
山口 栞 1751118: M, 1回目発表 知能コミュニケーション 中村 哲
title: Depressive tendency prediction in personal life data
abstract: Over 300 million people have suffered from depression all over the world. Although previous studies revealed some risk factors related to depressive symptoms, those studies did not take predictive performance into consideration. Theu did not use many variables to investigate either. This study examined the associations in 257 dimensions, and obtained high accuracy in random forest method. Future research will be comparison to previous studies, and reuction of the number of questions.
language of the presentation: Japanese