| 栢割 脩平 | M, 1回目発表 | ソフトウェア設計学 | 飯田 元, | 松本 健一, | 片平 真史（客員教授）, | 石濱 直樹（客員准教授）, | 高井 利憲（客員准教授） |
|
title: Uncertainty-Aware Deep Learning System for Autonomous Exploration Rovers
abstract: In recent years, improvement of the autonomy of the exploration rover has been required in order to efficiently explore a wider area in lunar and Mars exploration. However, in exploring unknown environments, conventional exploration rovers based on rule-based systems are limited in their exploration range. Therefore, in this research, the integration of an image recognition system based on deep learning into an exploration rover is investigated. In particular, a deep learning system architecture will be considered that incorporates two points: "uncertainty estimation method for prediction with low calculation cost" and "decision-making method based on uncertainty estimation value". By quantitatively evaluating the predictive uncertainty of the deep learning system and handing over control to the rule-based system when the uncertainty is high, we aim to achieve both high prediction accuracy and high safety. language of the presentation: Japanese 発表題目: 自律探査ローバーのための不確実性を考慮した深層学習システムの提案 発表概要: 近年、月・火星探査において、より広範囲を効率的に探査するため、探査ローバーの自律性の向上が求められている。しかし未知の環境を探査するうえで、ルールベースシステムに基づいた従来の探査ローバーでは探査範囲が限られてしまう。そこで本研究では、深層学習を用いた画像認識システムの探査ローバーへの搭載を検討する。特に、「計算コストの小さい予測の不確実性推定手法」と、「不確実性推定値に基づいた意思決定手法」の2点を検討し、それらを組み込んだ深層学習システムのアーキテクチャを提案する。深層学習システムの予測の不確実性を定量評価し、不確実性が高い場合はルールベースシステムに制御を引き渡すようにすることで、高い予測精度と高い安全性の両立を目指す。 | |||||||
| 川西 航太郎 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, | 中村 哲, | 舩冨 卓哉, | 藤村 友貴 | |
|
title: Automatic Colorization of Animated Line Drawings Using Multi Head Structure
abstract: In recent years, long working hours of animators have become a problem, and the automation of animation production has been attracting attention. One of the processes in animation production is colorization of line drawings. In the conventional automatic colorization method using deep learning, all characters to be colorized are learned by a common network. Therefore, there is a problem that the accuracy decreases when the number of characters increases. In this work, we propose an automatic colorization method using the Multi Head structure. In this method, the problem is simplified by coming down to attributing colorization to each character. Therefore, the probability that the accuracy is maintained even when the number of characters increases is high. language of the presentation: Japanese 発表題目: Multi Head 構造を用いたアニメーション線画の自動彩色 発表概要: 近年アニメーターの長時間労働が問題視されており、 アニメ制作の自動化が注目されている。アニメ制作の工程の一つとして線画の彩色が存在するが、従来の深層学習を用いた自動彩色では彩色する全てのキャラクターを共通のネットワークで学習しているため、キャラクターの数が増加した場合に精度が低下するという問題がある。そこで本研究では、Multi Head 構造を用いた自動彩色手法を提案する。本手法ではキャラクターごとの彩色に帰結させることで問題を単純化しているため、キャラクターの数が増加した場合も精度が維持されることが期待される。 | |||||||
| 星野 智紀 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, | 中村 哲, | 大内 啓樹 | ||
|
title:Contents controllable abstractive summarization
abstract:Recently, there has been a lot of research on abstract summarization. In these studies, there have been proposals for summarization models that can control the length and vocabularies of the generated summary sentences. These researches are expected to lead to further researches on summarization models that can control the content of the generated summary sentences. In this presentation, I will introduce the researches on controllable summary sentence generation and discuss the problems of these researches. language of the presentation:Japanese 発表題目:内容に関して制御可能な抽象型要約 発表概要:近年，抽象型要約の研究が盛んに行われている。それらの研究の中で生成される要約文の長さや含まれる語彙を制御できるような要約モデルの提案がなされている。これらの研究によって，今後，生成する要約文の内容に関して制御可能な要約モデルの研究が進むことが期待できる。本発表では制御可能な要約文生成に関する研究の紹介とそれらの研究の課題点について述べる。 slide | |||||||
| 永田 篤樹 | M, 1回目発表 | 知能システム制御 | 杉本 謙二, | 岡田 実, | 花田 研太 | ||
|
title: Feedback Error Learning Control against Signal Loss
abstract: Due to the progress of information and communication technology, networked control systems have been extensively studied recently. It is known that in networked control the sensing signal may be lost from time to time due to congestion in communication channels. Such temporal signal loss also happens due to occlusion of non-contact sensors. They cause serious problems for control. Hence we design switching output feedback controller, which guarantees the stability of the system under signal loss. Additionally, we propose 2DOF structure by using a Lyapunov solution derived from the FB control design. language of the presentation: Japanese 発表題目: 信号損失に対処するフィードバック誤差学習制御 発表概要: 近年，通信ネットワークや非接触センサを用いた制御が盛んに行われている．これらの制御手法ではパケットロスや外れ値などによる観測信号の損失が起こり，制御システムを不安定にする可能性がある．この問題に対処するために，信号損失下でもシステムを安定化させる切替型出力フィードバック制御器を設計する． また，この制御器を設計する段階で得られるリアプノフ解を用いることで，２自由度制御系への統合化設計が可能となる． | |||||||
| KIM GAHEE | M, 1回目発表 | 知能システム制御 | 杉本 謙二, | 池田 和司, | 松原 崇充 | ||
|
title: Parameter Inference of Black-boxed Simulators based on Likelihood-free Methods
abstract: In the robot learning field, it is common to use simulators to obtain training data. If we know a set of parameters that make simulators close to real environments,then the trained model can be easily transferred to the real world. However, since most simulators are implicit models, it is impossible to inference the posterior distribution of parameters from observations. In recent years, likelihood-free inference methods which approximate posterior using simulated data have been actively proposed to overcome this problem. In this research, we apply likelihood-free inference methods to the earth-work task using the Voltex studio simulator. language of the presentation: Japanese 発表題目: likelihood-free手法を用いたブラックボックスモデルのパラメータ推定 発表概要: ロボット研究分野ではシミュレータを使って学習データを集めることが一般的である。 実環境に近いシミュレータパラメータを設定することで、現実世界へのスムーズなモデル転移が可能であるが、 多くのシミュレータは、生成モデルが未知のブラックボックスモデルであることから観測結果からの逆算が不可能で パラメータの尤度関数を求めることができない。 近年、この問題を克服するために、シミュレーションでの観測データを元にニューラルネットワークを学習して データ中心的な方法で尤度関数を推定するlikelihood-free推定が数多く提案されている。 本研究では、これらの手法を応用し、土木シミュレータのvortex studioを用いて、現実世界の観測から 実環境を上手に表現するパラメータ推定を効率的に行う方法を検討する。 | |||||||
| 伊藤 和浩 | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 若宮 翔子, | 矢田 竣太郎 | |
|
title: Building a dataset of "complaints" and proposing a classification model
abstract: Complaining is an expression of negative words, and at the same time, it is an action that has a self-healing function. In this study, we construct a dataset of complaints based on the data obtained from Twitter, and use the dataset to build a machine learning model to classify whether a certain text is a complaint or not. In the future, the classifier can be used to improve the mental health of individuals and to improve the well-being of the workplace. language of the presentation: japanese 発表題目：「愚痴」のデータセット構築・分類モデルの構築 発表概要：愚痴はネガティブな言葉の表出であると同時に自己治癒の機能を持つ行為である．本研究ではTwitterから取得を行ったデータを元に愚痴のデータセットを構築し，そのデータセットを用いて，あるテキストが愚痴かどうかを分類する機械学習モデルを構築する．将来的には愚痴についての分類器を使って個人のメンタルヘルスの改善や，職場のwell-being向上につなげるなど，様々な応用が考えられる． | |||||||
| QU QIANYUE | M, 2回目発表 | 大規模システム管理 | 笠原 正治, | 岡田 実, | 笹部 昌弘, | 張 元玉 |
|
title: Millimeter Wave Vs. Microwave: Which Do Eavesdroppers Prefer?
abstract: Hybrid communication systems where sub-6 GHz links coexist with millimeter-wave (mmWave) links have become an essential combination in 5G networks. Nevertheless, the open feature of the hybrid systems makes them vulnerable to eavesdropping attacks. For eavesdroppers in hybrid communication networks, they may enhance their eavesdropping performance by selecting eavesdrop on different waves (i.e., mmWave and microwave). Hence the entire network can be divided into a mmWave eavesdropping region and a microwave eavesdropping region. We, therefore, investigate the eavesdropping region characterization problem in hybrid wireless systems from the perspective of physical layer security. First, we derive the lower bound of secrecy rate in mmWave link and secrecy rate of the microwave link, respectively. The simulation results verify the feasibility of the theoretical derivation. Second, we use the ratio between the secrecy rate of mmWave links and that of microwave links as the eavesdropping wave selection criterion to determine the mmWave and microwave eavesdropping region. Last, We provide several numerical results to illustrate the millimeter-wave eavesdropping area under various network parameter settings. language of the presentation: English | ||||||
| LE VU TRUNG DUONG | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 笠原 正治, | TRAN THI HONG(客員), | 張 任遠 |
|
title: *** MRSA: A High-Ef?ciency Multi ROMix Scrypt Accelerator for Cryptocurrency Mining and Data Security ***
abstract: *** Blockchain is a technology using a decentralized network. Accordingly, its database stores information differently from a typical database. The data is stored in blocks that are chained together. Once the block is filled with data, it will be stored in the blockchain permanently if proof of consensus mechanism is accepted. Blockchain is usually applied for cryptocurrencies such as Bitcoin. Bitcoin uses the Proof of Work(PoW) consensus mechanism. Despite ensuring high security, PoW required huge computation power to mine a new block to lengthen the blockchain. Moreover, because of the simple mining hash algorithm, most Bitcoin miners are performed by ASICs. With the outstanding hash performance, ASICs make the miners' power not balance and tend to concentrate. Then, the Bitcoin mining difficulty will increase, and ASIC miners will take over the mining process. This may harm the blockchain network because the ASIC miners control and make any invalidated data. Hence, ASIC-resistant algorithms are invented to destroy the ASIC advantages. They make ASICs useless, inefficient, or high cost and risk. With high complexity, they bring obstacles for general-purpose GPUs and CPUs. This research will develop an optimized hardware architecture for an ASIC-resistant mining system on FPGA, a flexible, high-performance, and low-power hardware platform. We will prove that our proposed system has much higher energy efficiency for cryptography mining systems by evaluating power consumption and hash rate. *** language of the presentation: *** English *** | ||||||
| STIRAPONGSASUTI SOPICHA | 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, | 門林 雄基, | 諏訪 博彦, | 松田 裕貴 |
|
title: A Method for Upload Decision of Smart Home Data Resilient to Re-identification Attack
abstract: Recent developments of smart devices and appliances have led smart homes to support elderly people. However, some smart services from a service provider(s) such as elderly monitoring, life logging, appliance control and so on may contain privacy-sensitive data which a malicious attacker can illegally access them for re-identification. In this research, a decision making method for data generated in smart homes and then uploaded to a cloud server(s) is proposed. The threat model is that an attacker is capable of monitoring some smart homes in a target area from outside in some time slots. Thus, there is a possibility that the attacker can re-identify data of dwellers by matching physical-observed event sets and cloud event sets. To solve this problem, I formulated the re-identification problem by considering types of data and upload frequency based on dweller budgets. Also, the multi objective optimization is selected to achieve upload decision while protecting data from re-identification. language of the presentation: *** English *** | ||||||