| 福本 晃汰 | M, 1回目発表 | 知能システム制御 | 杉本 謙二 | 池田 和司 | 小林 泰介 | ||
|
title: Towards Improving Computational Cost of Cross Entropy Method for ModelPredictive Control
abstract: In recent years, many of the systems and robots that are expected to be used in the real world are with multidimensional and nonlinear dynamics. Nonlinear model predictive control (MPC) has been actively studied as a means of safely and robustly controlling such systems. It has also received attention in the context of model-based reinforcement learning, where models are not given explicitly, but learning techniques infer them from data. However, the nonlinear MPC suffers from its computational cost caused by an iterative numerical solver. This problem would be fatal in systems that require control in real time. In this research, I focus on a cross entropy method (CEM), which is used for one of the nonlinear MPCs. For the CEM, I investigate methods to reduce the computational cost through numerical simulations. language of the presentation: Japanese | |||||||
| 武田 悠佑 | M, 1回目発表 | 知能システム制御 | 杉本 謙二 | 安本 慶一 | 小林 泰介 | ||
|
title: Adversarial Behavioral Cloning from Observation
abstract: Behavioral Cloning from Observation (BCO) enables robots to imitate experts' behaviors even without action sequence. However, this capability cannot be acquired unless the robots interact with environments while sacrificing sample efficiency. To improve sample efficiency, it is desirable to efficiently explore environments, especially in the early stages of learning. In this study, a disturber as a exploration policy is introduced to facilitate exploration. It is optimized through adversarial learning against the skill for inferring the experts' actions. The simulation results of the proposed method are reported. language of the presentation: Japanese | |||||||
| WIRAATMAJA CHRISTOPHER | M, 1回目発表 | 大規模システム管理 | 笠原 正治 | 松本 健一 | 笹部 昌弘 | 張 元玉 | |
|
title: Cost-Efficient Blockchain-Based Access Control
abstract: In these past few years, the number of Internet of Things (IoT) resources connected to the Internet keeps growing every year. Unfortunately, the growth of IoT resources number introduces a new security risk if left untreated. Current research trends try to mitigate this risk by developing Blockchain-Based Access Control. However, the cost of maintaining this type of Access Control is very costly due to the storage cost of Blockchain. This research finds a novel way to reduce Blockchain-Based Access Control storage cost without introducing a security risk by utilizing Decentralized Storage and Blockchain Oracle technology. language of the presentation: English | |||||||
| 坂井 優介 | 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: Adaptation of Knowledge Graph to Neural Machine Translation
abstract: Machine translation has been known to suffer from the problem of inadequate translation for unobserved words or low frequent words, such as named entities. In order to solve this problem, recent machine translation adapts subword units in order to expand the word vocabulary coverage while keeping the model's vocabulary size limited. However, merely splitting words into subwords does not solve the sparsity problem especially for newly created entities. Meanwhile, there exist rich human crafted resources, such as knowledge graphs, which might be incorporated into machine translation to directly solve the sparsity problem. However, given the huge graph size with enormous number of entities, it is not immediately clear how the resource could be integrated with machine translation. I propose two methods for injecting knowledge graphs directly into neural machine translation which do not affect the model's vocabulary size. In this presentation, I will introduce prior studies and discuss the future direction of my research. language of the presentation: Japanese | |||||||
| 中岡 黎 | 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 | 清川 清 | 諏訪 博彦 | 松田 裕貴 | 中村優吾 |
| 林 涼弥 | 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 | 藤本 まなと | 松田 裕貴 |
|
title: Estimating Emotion and Satisfaction of Tourists During Sightseeing Considering Weather Conditions
abstract: Smart tourism which provides rich tourism support to tourists using digital devices, is becoming increasingly common, but does necessarily not reflect the experiences of individual tourists. To provide richer tourism support, we have to recognize the psychological state of each tourist, such as emotional status and satisfaction level. In this study, we assume weather conditions affect their psychological states during sightseeing, and build a model for estimating tourists' emotion and satisfaction considering weather conditions. language of the presentation: Japanese 発表題目: 観光中の気象状況を考慮した観光客の感情・満足度推定 発表概要: デジタルデバイスを用いて観光客に質の高い観光支援を行うスマートツーリズムが誕生しつつあるが，こうした観光情報は必ずしも個々の観光客の体感を反映したものではない．より有益な観光情報を提供するには，個々の観光客の感情や満足度といった心理状態を認識する必要がある．本研究では，その心理状態がが観光中の気象状況にも影響されると仮定し，気象状況の影響を考慮した観光客の感情・満足度の推定モデル構築の手法を検討する． | |||||||
| 胡 尤佳 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 | 渡辺 太郎 | 須藤 克仁 | Sakriani Sakti | ||
|
title: Using Posterior Distribution as Intermediate in Multitask End-to-End Speech Translation
abstract: In this global world, speech translation (ST) system is useful when we have to communicate with foreign people in non-native language. ST aims to learn transformations from speech in the source language to the text in the target language. ST has been realized with a concatenation of automatic speech recognition (ASR) and machine translation (MT) system, which is called cascade approach. In these days, instead of cascade approach, End-to-End approach is becoming main approach in ST research. Previous works show that multitask learning improves End-to-End ST performance, in which the recognition decoder generates the text of the source language, and the translation decoder obtrains the final translations based on the output of the recognition decoder. However, in the general intermediate ASR task, the ambiguity of ASR output caused by acoustic similarity cannot be considered because it is calculating cross entropy loss with one-hot reference. In our research, we use ASR word posterior distribution as soft target loss when calculating sub ASR task loss instead of hard one-hot loss. Our proposed method improved BLEU score compared with cross entropy loss baseline in Japanese to English End-to-End Speech Translation task with BTEC corpus. language of the presentation: Japanese | ||||||||
| 土肥 康輔 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 | 渡辺 太郎 | 宮尾 知幸 | 須藤 克仁 | Sakriani Sakti | |
|
title: Improving the Performance of Grammatical Error Correction Models Using Pseudo Data Based on Error Analysis
abstract: Grammatical Error Correction (GEC) is the task of automatically correcting grammatical errors in text. Due to the lack of data for the task, incorporating pseudo data in the training of GEC models has been one of the main ways to improving the performance of the models. We analyze error types where existing GEC models perform poorly, and generate pseudo data related to the error patterns. By adding such data, we aim to improve the performance of GEC models. language of the presentation: Japanese | ||||||||
| 池田 聖華 | M, 1回目発表 | 数理情報学 | 池田 和司 | 松本 健一 | 久保 孝富 | 吉本 潤一郎 | 福嶋 誠 | 日永田智絵 |
|
title: Gaze Behavior Analysis
for Program Comprehension
abstract: Programmers spend much of their working time to comprehend programs. Therefore, efficient program comprehension is expected to improve the overall software development productivity. Previous research works have shown that programmers move their gaze strategically during reading programs. However, there has been no quantitative evaluation of programmers’ gaze, based on their importance of components of source code. Since it is expected that human attention is directed to important components, it might be able to assess programmers' gaze behavior by quantifying their importance. In this study, we aim to clarify the cognitive process of program comprehension by evaluating the importance of components in source codes. To this end, we compare programmers' gaze distribution with the machine learning model's attention as the importance of their programs. In the experiment, subjects performed a program categorization task and we measured their gaze during the task. We will investigate the relationship between the differences in performance and gaze distributions across subjects, and compare the similarity of machine learning models. In this presentation, I will report a preliminary result of an analysis in which we compared the gaze distributions of subjects. languege of the presentation: Japanese | ||||||||
| 石田 豊実 | M, 1回目発表 | 数理情報学 | 池田 和司 | 松本 健一 | 久保 孝富 | 吉本 潤一郎 | 福嶋 誠 | 日永田智絵 |
|
title: Elucidating the neural bases of program comprehension in expert programmers
abstract: Program comprehension requires much time and effort in software development. Previous studies aimed to understand strategies of expert programmers that lead to efficient program comprehension by analyzing their behavioral performance and physiological signals. However, the neural bases for program comprehension in expert programmers are still unclear. The purpose of my study is elucidating the neural bases of program comprehension by analyzing magnetic resonance imaging data of programmers obtained during a program categorization task. In my presentation, I will report basic results for the association between program comprehension and brain structure. language of the presentation: Japanese 発表題目: エキスパートプログラマのプログラム理解を支える神経基盤の解明 発表概要: ソフトウェア開発時に多くの時間と労力を要するプログラム理解を効率的に獲得するために，開発効率の高いプログラマ(熟練者)の特徴を把握しようとする試みがある． 既存研究では，熟練者とそうでない者でコーディングおよびデバッグのスピードや，視線移動・脳機能といった生体情報に違いがあることが報告されている． しかし，エキスパートプログラマのプログラム理解を支える神経基盤はいまだに明らかにされていない． 本研究ではプログラム理解と脳の構造・活動に関連があるかを調査する． 本発表ではプログラム理解と脳構造の関連についての基礎的検討結果について発表する． | ||||||||