| 新妻 巧朗 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, | 中村 哲, | 進藤 裕之 | |
|
title: Towards building the framework analyzing latent representations in Masked Language Model
abstract: Recently, BERTology, which is the study on Masked Language Model (MLM), is rising rapidly since MLMs have gotten great results in some NLP tasks. So, we want to use them on tasks a lot more, but they have some obstacles. For example, MLMs is likely to have many layers, so we must find out fitted layer on target tasks. I try to study to remove their obstacles and to build a common framework for the study method. In this study, I will analyze latent representations on each layer in BERT. language of the presentation: Japanese 発表題目: Masked Language Modelの潜在表現の分析フレームワーク確立に向けて 発表概要: 近年、Masked Language Model (MLM)に関する研究であるBERTologyが急速に興隆している。 これはBERTを端に発してMLMが、いくつかのNLPタスクにおいて大きな成果を残したためである。 そのため、もっと多くのタスクにこのようなMLMsを適用していくことが試みられている。 しかし、その際にいくつかの障害がある。 例えば、MLMは多くの層を持っていることが多いため、応用したいタスクに適切な層をヒューリスティックな方法で見つける必要がある。 ゴールとしては、このような障害を取り除くための研究とその研究のためのフレームワークを提案を目指したい。 本研究では、BERTのそれぞれの層の潜在表現を分析しようとしている。 | ||||||
| 山田 暉 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | 中村 哲, | 新保 仁, | 進藤 裕之 |
|
title: Sarcasm Detection using Contexts in Dialogues
abstract: Recently, sentiment analysis and emotion detection in dialogues are developing rapidly. However, these reseach have difficult problems since sarcasm expression often appears in Dialogues. Detectiong emotion in dialogue often fails due to sarcasm expression. In this research, we concentrate on sarcasm detection in order to solve this problem. We propose our model which detects sarcasm expression in current utterance using previous utterances. This reseach might help sentiment analysis and emotion detection. language of the presentation: Japanese | ||||||
| LI WAI HEI | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | 中村 哲, | 新保 仁, | 進藤 裕之 |
|
"title:Generation of Medical Report using Convolutional Neural Network and Long-Short Term Memory Neural Network"
"abstract: In this paper, we present a method to generate medical reports automatically using neural network. The possibility of generating medical report by computer is always an interest as the demand of medical report generation is high while the generation process itself is also repetitive. So, in this research, we will try to generate a medical report. Which medical images will be the input of the system, and the system will output a medical report which correctly describing the medical image under the format of existing reports. The system will also output the analysation of the medical images for reference. In this research, we will build the system will two distinct neural network, which are Convolutional Neural Network and Long-Short Term Memory Neural Network respectively, while utilizing the information from medical reports and medical images. We will demonstrate that the finding we have in the research will help the development of generating medical reports by computer, which ultimately we hope the process can be done without guidance from human." "language of the presentation: English" |
||||||
| 佐々木 皓大 | 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, | 中村 哲, | 諏訪 博彦（特任准教授） | |
|
title: Construction of trading simulation for evaluating investment risk prediction model
abstract: In recent years there has been an increase in the number of people who choose stock investment for asset management. When doing actual investment and asset management, avoiding risks is a very important issue. One of the indices used as a risk index is the Nikkei VI in japan. Suwa et al. (2017) analyzed stock bulletin board messages and predicted the rise of the Nikkei VI. In this study, I developed a simulation for trading Nikkei stock index option using intra-day data and verified the validity of the Nikkei VI prediction model proposed by Suwa et al. As a result, the improvement using their model was +6,611 yen and I confirmed Suwa et al.'s Nikkei VI prediction model is effective. Moreover, the expected goals and our current progress will also be reported here. language of the presentation: Japanese | ||||||
| 松井 智一 | 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, | 中村 哲, | 諏訪 博彦（特任准教授） | |
|
title: Easy-To-Deploy Living Activity Sensing System and Data Collection and Analysis in General Households
abstract: Emergence of smart appliances and high performance IoT devices is promoting studies on more functional and intelligent home services using these devices. In Japan, we are facing the problem of aging population and declining birthrate, hence it is urgent to develop technologies to improve resident’s QoL and monitor the elderly through home services based on the activity recognition technology. However, an activity recognition system in general requires many types/number of sensors and hence it is difficult to deploy and operate it in general households. In this paper, we propose a system consisting of low-cost and easy-to-deploy sensors that collects data of resident’s activities of daily living (ADL). The system was deployed in actual homes of senior citizens and collected ADL data for two months. We also estimated the ADLs from the collected data by Long short-term memory (LSTM), a deep learning model. As a result, ADLs could be estimated at high recall rate and hence we found that the proposed system has high applicability to actual services. language of the presentation: Japanese 発表題目: 一般家庭における居住者の日常生活行動センシングシステムの提案 発表概要: 家電のスマート化やIoT機器の高性能化を背景に，宅内サービスの高機能化が研究されている．特に，我が国では少子高齢化が進行していることから，生活行動推定技術を用いた宅内サービスによる居住者のQoL向上や，高齢者の見守りが切望されている．一方で，生活行動推定は多種・多数のセンサを要するため，一般家庭への設置・運用が難しいという課題がある．本研究では，安価かつ設置・運用が容易なセンサからなる生活データ収集システムを構築した．構築したシステムを一般の高齢者家庭に設置し，2カ月間の生活データ収集を行った．収集したデータを元に，LSTM（Long short-term memory）を用いた深層学習モデルにより居住者の行動推定を行った結果，高い再現率で居住者の行動推定を行うことができ，実サービスへの応用可能性を確認した． | ||||||
| ZHANG YAN | D, 中間発表 | 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 | ||
|
Title: Wide-view Optical See-through Head-mounted Displays with Per-pixel Occlusion Based on Conic Mirrors. Abstract: Optical see-through head-mounted displays (OSTHMDs) have been actively developed as a crucial platform for realizing augmented reality (AR), while the lack of mutual occlusion, which results in “ghost-like” images, and the narrow field of view (FOV) prevent existing OSTHMDs improving user experience further. With our work, an approach based on the optical structure of a conjugated conic mirrors pair that achieves a super wide FOV and the per-pixel occlusion capability in OSTHMDs is introduced. The system consists of a pair of ellipsoidal mirrors has been developed, a bench-top prototype has been built that a monocular FOV of H122° x V74° with per-pixel occlusion is demonstrated. However, the viewpoint offset and artifacts from the diffraction of the transmissive spatial light modulator (SLM) restrict the visual performance. So that a double parabolic mirror system is further proposed to address these issues, and also, challenge to expand the narrow FOV of the existing volumetric display into a naked-eye-like range. Language of the presentation: English | |||||||
| 内嶺 佑太 | M, 2回目発表 | インタラクティブメディア設計学 | 加藤 博一, | 清川 清, | 太田 淳, | 神原 誠之, | 藤本 雄一郎 |
|
title: Evaluations of Eye Rotation Tracking Application on HMD Using Micro-Lens Array
abstract: One drawback of the micro-lens array HMD, that can achieve a very small form-factor, is the image effects caused by eye-motion. Addressing this problem requires eye-tracking function that provides angle data of eye rotation for a real-time image calculation. In this study, a prototype of a micro-lens array HMD that comes with an eye-tracking function is built. Pre-evaluation results on the prototype have shown a positive result on the effectiveness of eye-tracking application, besides side-effects. Language of the presentation: English | |||||||