| 黒木 琢磨 | M, 1回目発表 | 数理情報学 | 池田 和司 |
|
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。 | |||
| 高阪 翔 | M, 1回目発表 | 数理情報学 | 池田 和司 |
|
title:Feature selection of nonlinear system using SVM
abstract:Neuronal axons are protrusions of nerve cells and are important structures for transmitting information to other nerve cells. When the spinal cord where many axons exist is damaged, the function of information transmission is lost, and it impairs motor function and sensory function. Therefore, the treatment of spinal cord injury and the development of drugs are underway. Cells function by nonlinear interaction by various molecules. Drugs acting on cells are generally compounds that inhibit the action of molecules, and which molecule group to inhibit is the key to drug discovery to control cell function. The purpose of this study is to select inhibitory molecules to further promote nerve axon elongation. The data set to be analyzed is the degree of elongation / regression of axons with respect to the inhibition pattern of 200 or more types of molecules. Therefore, a specific object is to extract an inhibition pattern in which the axon extends more from the combination of the scalar values of the axon outgrowth to the input vector of the inhibition pattern. For this purpose, we used the feature selection method (Koyama et al, ICONIP 2018) for the nonlinear data set of the previous study in this study. Koyama's method is superior to other feature selection methods (random forest importance, wrapper method, Relief method) in nonlinear dynamical systems. As a result of comparing these methods, it was shown that the feature selection method of Koyama et al. Is effective for the object of this research, that is, the network of biochemical reactions expressed by ordinary differential equations. language of the presentation:Japanese | |||
| 古志 将樹 | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清 |
|
title: ***Improvement of concentration by attenuating visual noise using video see-through HMD ***
abstract: ***When you want to concentrate on work, you may put earplugs or listen to music. It is because the learning efficiency can be improved by attenuating the auditory noise of the surrounding environment. As auditory noises, visual noises such as moving objects in the vicinity of work environments hinder the concentration of workers. In this paper, we propose a method to reduce visual noise around work environment and improve learning efficiency by putting HMD (Head Mounted Display) on the worker and adding effects of grayscale and blur out of working area. As a result, improvement of learning efficiency was confirmed. *** language of the presentation: *** Japanese *** | |||
| 斎藤 悠太 | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清 |
|
title: Presentation of difference in tactile dimensions by aerial encounter type haptic interface using drone
abstract: In the encounter type haptic interface in VR and AR, robot arms are mainly used, but by using drones, it is possible to realize an aerial encounter type haptic interface which enables the experienced person to move freely. Also, thinking that multiple tactile sensations can be presented by attaching multiple textures to the dron be available. In this study, we attached a plurality of textures of different hardness to the drone, fly the drones at a constant speed against the tip of the fixed bar, measure the amount of deformation of the texture at the time of contact with the RGB-D sensor It was confirmed that the deformation amount varies depending on the hardness. language of the presentation: Japanese | |||
| 阪本 充輝 | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸 |
|
title: Automated Segmentation of Hip and Thigh Muscles in Metal Artifact Contaminated CT using CNN
abstract: In total hip arthroplasty, analysis of postoperative images is important to evaluate surgical outcome and create appropriate rehabilitation plans. The challenge we addressed in this work is the metal artifact in postoperative CT caused by the metallic implant, which reduces the accuracy of segmentation especially in the regions next to the implant. Our goal is to develop an automated muscle segmentation in the postoperative CT images. In this research, we propose a method that combines Normalized Metal Artifact Reduction (NMAR), which is a state-of-the-art metal artifact reduction method with a CNN- based segmentation using the U-Net architecture. We conducted experiments using simulated images and real images of the lower extremity to evaluate the segmentation accuracy of 19 muscles that are contaminated with metallic artifact. In simulation study, the proposed method improved the average symmetric surface distance (ASD) from 1.85 ± 1.63 mm to 1.24 ± 0.67 mm (mean ± std). The real image study using two CT images with the ground truth of gluteus maximus, medius and minimus muscles showed the reduction of ASD from 1.67 ± 0.40 mm to 1.52 ± 0.47 mm. Our future work includes the development of an end-to-end convolutional neural network for metal artifact reduction and musculoskeltal segmentation and to establish a ground truth dataset by performing non-rigid registration between the postoperative and preoperative CT of the same patient. language of the presentation: Japanese | |||
| 佐藤 開 | M, 1回目発表 | 数理情報学 | 池田 和司 |
|
title: *** Investigation of SPLICE for brain information expression learning method by Micro-state ***
abstract: *** The number of people suffering from medical institutions due to mental illness is increasing.Research using brain information is advancing in the field of diagnosis and treatment of mental disorders.Research on fMRI is advancing, but diagnosis is expensive.Although the reliability of brain state estimation due to electroencephalogram is considered low, it is expected as an inexpensive diagnosis and treatment method.It is suggested that SPLICE and Micro-state method can distinguish between depression and healthy subjects.The objective of this research is to connect to SPLICE diagnostic and therapeutic methods of psychiatric disorders by applyingMicro-state and acquiring new findings.Currently, I am trying to master SPLICE analysis method.In the future, we analyze simultaneous measurement data of EEG and fMRI and examine what kind of trend can be seen. *** language of the presentation: *** Japanese *** | |||
| 佐野 友哉 | 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 Scheduling System for Picking Up Elderly People Reflecting Affinities among Passengers
abstract:In recent years, due to the aging of the population and the depopulation of rural areas, the shortage of transportation for elderly people has become a problem. In order to solve this problem, local governments and local volunteer staff are offering pick-up service, but the management of the service is hand-powered, which is troublesome and costly. The service managed automatically are also available, but in many cases the collection of pick-up requests is performed by the application. Most elderly people have trouble sending the demand because they are unfamiliar with a complicated operation. Therefore, in this research, we propose a system that collects the pickup requests using devices that are easy for the elderly to operate and automatically makes a schedule to pick up. If passengers who do not get along with each other share a car, the troubles can be caused. Therefore, when making a pick-up schedule, the system considers the affinities among passengers to improve comfort during their riding. language of the presentation:Japanese | |||
| 清水 航 | M, 1回目発表 | ソフトウェア設計学 | 飯田 元 |
| 白井 侯丞 | M, 1回目発表 | ネットワークシステム学 | 岡田 実 |
|
title:Wireless position location over indoor multipath channel using LCX-MIMO system
abstract: In our laboratory, a Multiple-Input Multiple-Output (MIMO) system using a leaky coaxial cable (LCX) has been proposed as a method of wireless position location in indoor. This system enables us to provide not only broadband communication, but also wireless position location for employing an LCX as an antenna. In previous studies, its precision has compared between two methods based on Multiple Signal Classification (MUSIC) and Estimation of Signal Parameters via Rotational Invariance Technique (ESPRIT) over the channel measured in anechoic chamber. However, verification on a fading channel including reflection and diffraction has not been done so far. Therefore, in this study, its precision will be evaluated over multipath fading channel through computer simulation. language of the presentation: Japanese | |||
| 新谷 隆太 | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 |
|
title: A Compression Method for Amount of Transfered Data on Highly Efficient Inference Execution Model with CNN
abstract: Since the amount of computation required for inference processing of the neural network is large, it is processed not by the location getting input data but by a server with abundant calculation capability installed in the data center. However, in overconcentration processing at the server, enormous power consumption and traffic congestion occur due to data transfer. Therefore, assuming the inference processing of the neural network model in edge computing to be implemented in the future, it is possible to alleviate traffic congestion that constructing an efficient distributed implementation model of Deep Neural Network. In the supposed model, the amount of data to be transferred is reduced by executing some processing on the edge computing rather than directly sending the input image. Furthermore, it is possible to reduce the amount of transfered data by compressing the feature amount of intermediate data. In this presentation, we propose a method to classify intermediate data into two classes of feature quantity channel and channel with few feature quantity channel per each batch number. Consequently, it increases continuous of equivalence value. language of the presentation: Japanese 発表題目: CNNを用いた高効率な推論実行モデル上の転送情報量に対する圧縮手法の提案 発表概要: ニューラルネットワークの推論処理に必要な計算量は大きいため、データ取得場所付近ではなく、データセンタに設置された潤沢な計算能力をもつサーバで処理される。しかし、サーバにおける一極集中処理は、膨大な消費電力やデータ転送によるトラフィック輻輳が問題となる。そこで、将来的に実装されるエッジコンピューティングにおけるニューラルネットワークモデルの推論実行を想定し、Deep Neural Networkの効率的な分散実装モデルを構築することで問題であるトラフィック輻輳の緩和が可能となる。想定するモデルにおいて、入力画像をそのまま送るよりもある程度の推論処理をエッジ側で行うことで転送するデータ量が削減される。さらに、転送する中間データの特徴量を圧縮することで転送情報量を削減が可能となる。本発表では、中間データをバッチ数毎に特徴量の多いchannelと特徴量の少ないchannelの2クラスに分類し、同値の連続性を増加させる手法を提案する | |||
| 鶴山 優季子 | 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: Improving accuracy of rent estimation for restaurants using objective and latent information
abstract: Along with the development machine learning, rent estimation is widely done. However for restaurants, rent estimation is determined based on tacit knowledge of veteran salesman such as experience or intuition, the price is different from person to person and it’s difficult for rookies to learn tacit knowledge. Therefore, in previous research, they proposed a rent estimation model for restaurants based on static, dynamic, and latent information. However it does not have sufficient accuracy. Therefore in this research, we try to improve accuracy by efficient extracting feature values and creating model. In particular, we add efficient feature values, establish the methods of extracting objective indicator, and extract latent information using new natural language processing. language of the presentation: Japanese 発表題目: 客観的・潜在的情報を用いた飲食店向け不動産の賃料推定精度の向上 発表概要: 機械学習の発展に伴い，不動産の賃料推定は広く行われている．しかし飲食店向け不動産ではベテラン営業マンの経験や勘といった暗黙知に基づいて賃料の価格が決定されているため，不動産価格が人により異なる，また新人職員への知識継承が難しいといった問題がある．よって先行研究では，飲食店向け不動産の静的情報，動的情報，潜在的情報に基づいた賃料推定モデルを提案した．しかし，実運用で用いるには精度が不足している．そこで本研究では，より効果的な特徴量の抽出およびモデルの構築により精度向上を試みる．具体的には，インタビューに基づく効果的な特徴量の追加，客観的指標の抽出手法の構築，新たな自然言語処理技術を用いた潜在的情報の抽出を行う． | |||
| 曽根田 悠介 | 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 |
|
title: Investigation of Quality Measurement System for Online Meeting
abstract: In Japan, the government has prepared an enviroment to work remotely. But there are not many people using remote video conference systems such as Skype. In this paper, I propose a method for measuring engagement of people participating in a meeting for online meetings. language of the presentation: Japanese 発表題目: オンラインミーティングを対象とした会議の質評価システムの検討 発表概要: 働き方改革などをはじめ，様々な場面でリモートで働く環境を整える機会が増加している． しかし，Skypeといったリモートでのテレビ電話会議のシステムを利用している人はまだ多くないというのが現状である． 当研究では，オンラインミーティングを対象とし，会議に参加している人の表情などを解析しミーティングの質の評価を計測する手法を提案する. | |||
| 田井中 渓志 | M, 1回目発表 | インタラクティブメディア設計学 | 加藤 博一 |
|
title: The construction and the effectiveness verification of the AR content production guideline for work supporting.
abstract: Currently, our laboratory creates AR content production guidelines for work support. We need to verify this because it is not guaranteed whether the application created by this guideline works properly. At the same time, if there is a problem with the guideline, we will improve it based on the verification result. language of the presentation: Japanese 発表題目: 作業支援のためのARコンテンツ製作ガイドラインの構築と有効性の検証 発表概要: 現在、我々の研究室では、作業支援のためのARコンテンツ製作ガイドラインを作成している。このガイドラインで作成したアプリケーションが適切に機能するものなのかの検証がされていないので、この検証を行う。それと同時にガイドラインに問題があれば、検証結果をもとに改善を行う。 | |||
| 高須賀 昌烈 | M, 1回目発表 | サイバーレジリエンス構成学 | 門林 雄基 |
|
title: Detecting C2 communications using network covert channels
abstract: Network covert channels are methods of data transfer unintentional using with network protocols and specifications. Encryption converts original texts into another texts, thus protects the interception from adversaries. On the other hand, network covert channels can transfar malicious texts because it disguise as a part of ligitimate communnications. The existence of encryption communications can detect but the existence of communications using network covert channel can not detect easily. Nowadays, some malware creaters use this methods as command communications of malwares. This methods are expected to use as the hiding methods of transfaring the exploited information from target companies. The goal is the detection of C2 communication using network convert channel and I consider exsting methods of network covert channels. language of the presentation: Japanese | |||
| 髙田 大樹 | M, 1回目発表 | ソフトウェア工学 | 松本 健一 |
|
title: Improving Algorithm Implementations in Software Using Online Judge Datasets
abstract: Online judges, also known as competitive programming, are online environments in which testing takes place for source code that is written to solve given problems. To solve problems with limited execution times and other constraints, effective algorithms are implemented, hence online judge repositories can be considered as knowledge resources of state-of-the-art algorithm implementations. Software development, however, may not adopt such implementations because of unawareness of recent algorithms. Because of different aims and characteristics, there is a big gap between software development and online judges, and then knowledge sharing has not occurred efficiently. In this research, we try to bridge the gap by applying effective implementations in online judges to software development. language of the presentation: Japanese | |||
| 隆辻 秀和 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 |
|
title: Individual neural conversation model based on a mixture of attributes
abstract: To realize human-like conversation is a long-term goal of dialogue systems. The Neural conversation model obtained the ability of more natural response generation. However, theirs remains several problems about generating responses. One of these problems is that generating response has lack of consistency. To deal this problem, We introduce new information, which cluster belongs to, to end-to-end conversation model. Our approach based on one assumption, consistency and wording style of a person is represented each cluster they belong to. Now, we built baseline-model and analyze it's generating results. language of the presentation: Japanese | |||
| 髙橋 洸丞 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 |
|
title: Analysis on A Metric of Human Judgement Score and Automatic Evaluation Metrics for Machine Translation
abstract: Human judgement scores are necessary for evaluating machine translation systems and automatic evaluation metrics. For the first step to build an automatic evalution system, we investigated the reliability of human judgement scores. Also, with that knowledge, we analized correlations betweeen human judgemnt scores and automatic evaluation metrics scores by comparing them directly. language of the presentation: Japanese | |||
| 辰巳 守祐 | M, 1回目発表 | 自然言語処理学 | 松本 裕治 |
|
title: Unsupervised Named Entity Recognition
abstract: We propose unsupervised named entity recognition model. We use two approaches. First one is subword contextual language model which reflects context information in word expressions. Secondly, we use Distant Supervision for creating a large amount of annotation corpus from raw corpus. In the end, we evaluate the model recognition rate of named entity not in dictionary. language of the presentation: Japanese | |||