ゼミナール発表

日時: 9月25日(火)1限 (09:20-10:50)


会場: L1

司会: 浦西 友樹
小澤拓 1151031: M, 2回目発表 加藤博一, 萩田紀博, 山本豪志朗, 浦西友樹
title: A Study on Suit Place of Projection with a Depth Camera and a Projector
abstract: Several research works have focused on a smart environment, an office with various kinds of sensors and displaying devices. On the other hand, group works have some discommunications because of a gap among attendees' knowledge. In this research, I propose a method for displaying information onto proper location by using a depth sensor and a projector. This presentation shows that a prototype of the proposed method.
language of the presentation: Japanese
発表題目: 距離画像センサ-プロジェクタを利用した適所投影システムの検討
発表概要: 近年、オフィスなどのいたるところにセンサやモニタなどの入出力機器を設置したスマート環境というものが注目されている。一方、グループワークにおいては参加者の知識の差や考え方の差によりコミュニケーションがうまく行われず、時間損失や機会損失に繋がることがある。スマート環境の情報提示機器により、グループワークに知識の差や、考え方の差を埋めるための情報の提供を行うことで、より高密度な議論を誘発し、時間損失や機会損失を防げるのではないかと考えた。
 
久保和樹 1151044: M, 2回目発表 加藤博一, 萩田紀博, 山本豪志朗, 宮崎純
title:A Window Arrangement Control in a Ubiquitous Display Environment
abstract:The display environment which used space widely can be built by combining two or more display apparatus. Such environment is expected as a place which performs collaboration work, because it can show much information at the same time. In such environments, there are some spaces of display which are not suitable for watching the window. In addition, the operation of the window take much time. Those points by the placement of the apparatus and the vastness of the environment can be suggested as present problems.Therefore, we suggest a function to control window arrangement automatically in consideration of display apparatus's placement of the real world. In this presentation, I will talk about the prototype system and the consideration result as my progress.
language of the presentation: Japanese
発表題目: 広大なディスプレイ環境におけるウィンドウ配置制御
発表概要: 複数のディスプレイ機器を組み合わせることで空間を広く利用したディスプレイ環境を構築することができる.そのような環境は多くの情報をウィンドウ形式で同時に並べて表示することが可能であり,共同作用を行う場として期待できる.現状の問題として,ディスプレイの配置やその広大さによってウィンドウを閲覧に適さない箇所が生じることやウィンドウの操作に手間が掛かることが挙げられる.そこで本研究では,実世界のディスプレイ配置を考慮して自動的にウィンドウ配置を制御する機能を提案する.本発表では,提案機能を有するシステムを試作し,考察を行った結果を進捗として述べる.
 
高橋達 1151061: M, 2回目発表 萩田紀博, 加藤博一, 神原誠之
title: A Social Media Mediation Robot to Increase Opportunity of Conversation for Elderly
abstract: This presentation introduces a social media mediation robot for increasing conversation opportunity of elderly. The some kinds of social media have been used by many users who can use digital device, such as PC or a smart phone. Though the opportunities of communication of these users are increasing, social isolation of elderly by digital divide is aggravated. In order to overcome this problem, we develop the social media mediation system by using a robot as interface to social media. We show the usefulness of our system through a pilot study to increase an opportunity for conversation.
language of the presentation: Japanese
 
守口裕介 1151107: M, 2回目発表 萩田紀博, 加藤博一, 浮田宗伯
title: grouping people for tracking between non-overlapping cameras
abstract: People tracking in a wide area can provide us useful information for various applications such as surveillance and navigation. In practical scenarios, wide-area tracking requires people identification between non-overlapping fields-of-view of distributed cameras due to cost concerns. People identification between the non-overlapping cameras is much difficult. This is because the appearance of each person is changed significantly due to changes in viewpoint and camera properties. For improving the robustness of this identification, in this work, groups of people are estimated in each camera and then utilized as an additional clue for people identification between different cameras; if two persons are in the same group in a camera view, they should be in the same group also in another camera view. The grouping process is achieved by discriminatively classifying the trajectories of grouped people and non-grouped people. In this talk, several experimental results for grouping people would be introduced.
language of the presentation: Japanese
 

会場: L2

司会: 小町 守
中清行 0951204: M, 2回目発表 鹿野清宏, 松本裕治, 猿渡洋, 川波弘道
title: Investigation on response accuracy of a speech-oriented information guidance system in a noisy environment with respect to topic limitation of the Question-Answer Database.
abstract: Spoken dialogue systems that run on mobile devices have been increased. As mobile devices are used in various environments, response accuracy decreases due to noise. If spoken sentence can't be recognized accurately, response accuracy decrease. Therefore technology that avoids the degradation of response accuracy in environments where recognition accuracy decreases is necessary. In this study, we propose a method to increase an accuracy rate, by limiting the number of topics in such environments. It was achieved by changing the size of question answer database(QADB). The proposed method will be evaluated using speech data recorded by a mobile phone by playing real environment data collected by Takemaru-kun system and additional noise together.
language of the presentation:Japanese
発表題目:音声情報案内システムにおけるデータベースのトピック制限による応答正解率の変化に関する調査
発表概要:携帯端末上で動作する音声対話システムが増加してきている.携帯端末は様々な環境下で利用されるため,雑音が多く認識精度が下がり,応答文の選択過程へ影響するため,応答精度が下がる.よって,認識精度が劣化する環境での運用でも応答正解率を下げないための技術が必要と考えられる. そこで本研究では,そのような環境下でトピック数を制限することで,全体の正解率を向上させる方法を提案する。トピック数の制限を質問応答データベース(QADB)の大きさを変化させることで実現し,それをたけまるくんのシステムとデータを用いて検討を行う.
 
真嶋温佳 1151097: M, 2回目発表 鹿野清宏, 松本裕治, 猿渡洋, 川波弘道
title: Invalid Input Rejection Using Bag-of-Words for Speech-Oriented Guidance System
abstract: On a real environment speech-oriented guidance system, a valid and invalid input discrimination is important as invalid inputs such as noise, laugh, cough and utterances between users lead to unpredictable system responses. Previously, acoustic features such as MFCC (Mel-Frequency Cepstral Coefficient) are used for discrimination. Comparing acoustic likelihoods of GMMs (Gaussian Mixture Models) from speech data and noise data is one of the typical methods. In addition to that, using linguistic features, such as speech recognition result, is considered to improve discrimination accuracy as it reflects the task-domain of invalid inputs and meaningless recognition results from noise inputs. In our work, we introduce Bag-of-Words (BOW) as a feature to discriminate between valid and invalid inputs. Support Vector Machine (SVM) and Maximum Entropy method (ME) are also employed to realize robust classification. We experimented the methods using real environment data obtained from the guidance system “Takemaru-kun.”
language of the presentation: Japanese
 
吉田雄太 1151116: M, 2回目発表 鹿野清宏, 松本裕治, 猿渡洋, 川波弘道
title: Examination on Topic Introducing Language Models for ASR
abstract: In recent years, practicalizing of speech recognition services such as Voice Search has been advanced. These devices use an n-gram model and it is the mainstream. An n-gram is modelized from the dependencies between short-range words, but an n-gram model does not take into account the dependencies of a wider range including context or topic. Then, with the aim of building language models introducing topic for ASR, this report examines about language models built from utterances collected by speech dialog system "Takemaru-kun".
language of the presentation: Japanese
発表題目: 自動音声認識のためのトピックを使用した言語モデルの調査
発表概要: 近年,Voice Searchなど音声をインターフェースとした情報機器の実用化が進んできた.それらのデバイスにはn-gramモデルが用いられており,現在主流となっている.n-gramモデルとは近距離の単語の依存関係をモデル化したものであるが,n-gramモデルは分脈情報やトピックなど,より広範囲の依存関係を考慮していない.そこで本研究では,トピックを導入した音声認識言語モデルの構築を目的とし,音声対話システム「たけまるくん」により収集された発話から構築される言語モデルについて調査を行った.
 
KANASHIRO PEREIRA LIS WEIJI 1151128: M, 2回目発表 松本裕治, 鹿野清宏, 新保仁, Kevin Duh
title: Collocation Suggestion System for Japanese Second Language Learners
abstract: Word combination, or collocation in a language is one difficulty found in second language acquisition. It is one of the common mistakes made by language learners. This study finds the best combination of words that can aid language learners, specifically, Japanese second language learners. We analyzed correct word combinations using different collocation measures and thesaurus based word similarity. The data are then evaluated against a Japanese Language Learner Corpus (Lang-8). Results show that using solely association measures gives better result than combining association measure and thesaurus based word similarity.
language of the presentation: English
 

会場: L3

司会: 久保 孝富
井林雅樹 1151014: M, 2回目発表 杉本謙二, 池田和司, 平田健太郎, 松原崇充
title: Learning a Kernel Matrix for Time Series Data from CTW Distance and Its Application for EMG Pattern Recognition
abstract: We propose a method for constructing a robust motion classifier from electromyogram (EMG) time series data against the user's postural changes that may result in largely different data even for the same motion. Our approach learns a kernel matrix from EMG time series data based on the canonical time warping (CTW) distance. CTW distance for two time series data is defined as the minimum distance of them through the optimal spatiotemporal alignment. In order to satisfy the positive semi-definiteness of the obtained kernel matrix, we use semi-definite programming. The motion classifier is then constructed with the obtained kernel matrix by a multi-class support vector machine. The experimental results suggest that the proposed approach is enable to have higher classification performance than a comparison method based on dynamic time warping (DTW) distance.
language of the presentation: Japanese
発表題目: 正準時間伸縮距離に基づくカーネル行列学習とEMG時系列パターン認識への応用
発表概要: 同一動作に対する時系列EMG信号を多様に変化させるユーザの姿勢変化に頑健なパターン識別器の構成法を提案する.提案法では,ペア時系列間の距離を最小とする時空間伸縮のもとで定義される正準時間伸縮距離に基づいてカーネル行列を学習する.そして,得られたカーネル行列を用いてマルチクラスSVMを学習する.手の4つの動作を識別対象とし,前腕より計測されたEMG信号を用いた評価実験の結果,提案法は動的時間伸縮距離に基づく比較手法よりも高い識別精度を示した.
 
田中大介 1151065: M, 2回目発表 杉本謙二, 池田和司, 平田健太郎, 松原崇充
title: Nonlinear System Identification with Input-Output Manifold Learning
abstract: This study considers a system identification scheme using high dimensional input-output data, e.g. images, spectra of sounds. Generally, the high dimensional data make identified system inaccurate. This may be due to the curse of dimensionality. In order to avoid this curse, a nonlinear system identification scheme with input-output manifold learning scheme is proposed. The input-output manifold learning is based on a nonlinear dimensionality reduction method regularized by the latent dynamics structure. The effectiveness of the proposed scheme is illustrated using artificial data.
language of the presentation: Japanese
発表題目: 入出力多様体学習を用いた非線形システム同定
発表概要: 本研究では, 画像や音の周波数スペクトル等の高次元入出力データを用いたシステム同定法について議論する. 一般的に高次元入出力データを用いた同定は次元の呪いにより困難である. これを回避するため, 入出力多様体学習を用いた非線形システム同定法を提案する. この手法は隠れダイナミクスの構造を用いて正則化された非線形次元削減法に基づく. この同定法の有効性を人工データにより示す.
 
安並健太郎 1151109: M, 2回目発表 杉本謙二, 池田和司, 平田健太郎, 松原崇充
title: Reinforcement Learning based on Iterative Optimal Control with Path Integrals
abstract: Stochastic optimal control with path integrals, an optimal control method by sampling, has some characteristic behavior. However, it can't be applied to some systems because of an equality constraint. As the solution for this problem, iterative stochastic optimal control with path integrals (iterative PI) had been proposed, but it's unimplemented and doesn't be demonstrated the validity and usefulness. In this study, a implementation method and reinforcement learning algorithm with iterative PI are proposed.
language of the presentation: Japanese
発表題目: 反復型PI法の有効性の検証と強化学習への応用
発表概要:非線形システムに対する最適制御則の一つである軌道積分確率最適制御法(PI)は,サンプリングに基づく最適制御則であり,他の既存法にはない特徴を持つが, アルゴリズムの導出過程で設定された等式制約によって扱えるクラスが制限されているという欠点がある. この問題を解決する方法として,反復型のPIが提案されているが,実装アルゴリズムやその有効性の検証については行われていなかった. 本研究では反復型PIの実装アルゴリズムを作成し,有効性を検証するとともに, PIの強化学習応用であるPI^2(Policy Improvement with Path Integrals)の考え方を用い,反復型PIを基にした強化学習アルゴリズムを提案する.