| 安形 俊輝 | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 |
|
title: Classification of Chest X-ray Images using CNN
Since it has been alarmed that a shortage of trained radiologist would affect on the patient treatment near future in Japan, computer aided diagnosis (CAD) is one of the most expected tool to improve the performance of doctors. In this study, we address to develop a model that detect abnormality in chest X-ray images using Convolutional Neural Networks (CNN). Though there has been various approaches of CAD based on CNN such as Variational AutoEncoder, most of them using a limited size of images as an input, so that it was difficult to identify detail texture features. We will introduce a two-stage CNN model that firstly reduces the search area by training an attention network to optimize a region of interest in order to process a higher resolution images and extract graphical features from the focused regions. language of the presentation: Japanese | |||
| 石坂 守 | M, 1回目発表 | ディペンダブルシステム学 | 井上 美智子 |
|
title: Area-Efficient and Reliable Hybrid CMOS/Memristor ECC Circuit for ReRAM Storage
abstract: Resistive random access memory (ReRAM) has several attractive features such as high-storage-density and highswitching with low power consumption. It is hence regarded as the most promising nonvolatile memory material. However, a memristor, which is a primitive component of the ReRAM-based memory, has much lower write endurance than that of dynamic random access memory (DRAM) or static random access memory (SRAM). Hence, error correction code (ECC) circuit is indispensable for realizing reliable ReRAM storage. For this purpose, we propose a hybrid CMOS/memristor-based ECC circuit. In the proposed circuit, the blocks with highly frequent write operations are implemented by the conventional CMOS technology and the others are implemented by the memristors to keep a balance between the area overhead and reliability. Through numerical experiments, we demonstrate that the proposed ECC circuit achieves less area while preserving the reliability compare to ECC circuits that are fully implemented by CMOS technology, where an area reduction of 19.9% is achieved for data words with 1,024 bits and the area reduction is improved as the data bit length increases. language of the presentation: Japanese | |||
| 浅井 沙良 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 |
|
title: Persuasive dialog system using multimodal emotional expression based on dialog management
abstract: Emotioal expressions are known as an efficient way to convey humans’ intents in conversions. On behalf of humans, a persuasive dialogue system using emotional expressions are studied, however, conventional studies have difficulties in generating diverse responses and directly managing response selection for persuasion because the systems search for highly similar utterances to humans’ ones. In this research, we aim to improve the efficiency of persuasion, and propose a dialogue agent which efficiently responds with multi-modal emotional expressions. This presentation provides you with an introduction of persuasive dialogue generators reflecting emotions. language of the presentation: Japanese 発表題目: 対話制御に基づくマルチモーダル感情表現を用いた説得対話システム 発表概要: 感情表現は、人間同士の対話において自分の意見を相手に受諾させる手段として効果的であることが知られている。人の代わりにユーザを説得する対話システム研究においても、感情表現を用いることの有用性が期待されているが、従来のシステムでは限られたテキストによる応答しか考慮されていないという問題や、ユーザ発話と類似度の高い発話文に対する応答文を検索していたため、説得を受け入れてもらうための制御を直接していないという問題があった。本研究では説得の成功率の向上を目指し、マルチモーダルな感情表現を用いて効果的に応答を行う対話エージェントを提案する。本発表では主に感情を反映させた説得対話の応答生成について紹介する。 | |||
| 石橋 陽一 | M, 1回目発表 | 知能コミュニケーション | 中村 哲 |
|
title: Encoding Sentences Using Logical Embedding
abstract: In the encoder-decoder, there is a problem that there is no semantic consistency in the encoded distributed representation. (Problem of language understanding) Furthermore, it is difficult for human to understand distributed representation that is outputed by the encoder. (Problem of visualization) In this research, we propose a encoding method that guarantees semantic consistency by fusion of symbol logic and neural networks in order to solve this problem. We expect to be able to encode different style sentence (but semantics is same) to same distributed representation. Moreover, By using the method, we can visualize a sentence as a logical formula. In this presentation, I describe the proposed method and the results of preliminary experiments. language of the presentation: Japanese 発表題目: 論理的埋め込みを用いた文の符号化 発表概要: encoder-decoderにおいて、符号化された分散表現には意味論的な一貫性がないという問題点がある（言語理解の問題）。 また、人間にとって、encoderが出力した分散表現を解釈することは難しい（可視化の問題）。 そこで本研究では、この問題の解決を目指すため、記号論理とニューラルネットを融合することで意味論的な一貫性を保証した符号化を行う手法を提案する。 提案手法により、どのようなスタイルの文であっても、意味論的に同じ文であれば、同じ分散表現として符号化できることが期待される。 また、文を命題論理式として可視化でき、解釈することが可能になる。 本発表では、提案手法のアルゴリズムと、予備実験の結果について述べる。 | |||
| 池澤 隼人 | M, 1回目発表 | 光メディアインタフェース | 向川 康博 |
|
title: Correspondence estimation and interpolation among line drawings using graph structure
abstract: Some software have been released with the function to assist animation production by interpolating line drawings. However, there are some problems such as the users are required to associate the strokes among frames to use this function and the connection of strokes are not guaranteed in interpolated results. For these reasons, the function is rarely used in actual animation production sites. Focusing on maintaining the connection relation between strokes in adjacent frames, I try to estimate the correspondence and interpolate given line drawings in this study. In particular, I will construct the graph structure from a line drawing. The nodes are the intersections of strokes and the edges are the sub-strokes between the nodes. I will develop algorithms to estimate strokes correspondence using graph matching and to interpolate line drawings under the condition that the graph structure is maintained. language of the presentation: Japanese 発表題目: ストロークの接続関係を考慮した線画間の対応付けの推定および線画補間 発表概要: アニメ制作の補助を目的として，線画を自動補間する機能を持った2Dアニメーションソフトウェアがある．しかし，これらの機能を利用するにはフレーム間でストロークの対応付けを行う必要がある上，補間結果でストロークの接続関係が保証されない問題があり，実際の現場での利用はまだ限定的である．隣接するフレームではストローク同士の接続関係がある程度維持されることに着目し，本研究ではこれを利用した対応関係の推定や補間の実現に取り組む．具体的には，線画に含まれるストロークの交点をノード，交点間のストロークをエッジとしたグラフを構築し，グラフマッチングによる自動対応付けおよびグラフ構造を維持する拘束を設けた線画補間を検討する． | |||
| 石井 大地 | M, 1回目発表 | 光メディアインタフェース | 向川 康博 |
|
title: Automatic colorization of Anime by cGAN and post-processing
abstract: Colorization task in the Anime production process still be forced manual colorization. It is time consuming task and wastes human resources. In this research, we realize an automatic colorization by cGAN with Anime dataset that is actually used at an animation production. In order to improve the colorization result, we apply post-processing. In this presentation, as the first step, we will report the result of automatic colorization of single character. language of the presentation: Japanese | |||
| 今岡 一章 | M, 1回目発表 | ネットワークシステム学 | 岡田 実 |
|
title: A Statistical Analysis of Relationship between Rainfall and GNSS Zenith Total Delay
abstract: As one of the meteorological observation techniques, the Global Navigation Satellite System (GNSS) is known. This technique utilizes the fact that the GNSS signal is delayed when it goes through troposphere with high water vapor in atmosphere, and causes positioning error. This extra delay time is treated as zenith total delay (ZTD), and the ZTD is internally estimated on positioning calculation in GNSS receiver. Although the ZTD has the local maximum before precipitation, it is not always satisfied. This study statistically analyzes relationship between precipitation and ZTD increment, and investigates the feasibility of rainfall prediction. Moreover, required observation period in time will be verified for detecting local heavy rain. language of the presentation: Japanese | |||
| 岩本 淳 | 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: Multiple Loop Correspondence Method of Linear Array with Intra-unit Feedback
abstract: In recent years, performance of Machine Learning Algorithm such as Deep Learning etc. greatly improves due to performance improvement of Computer. Demand for implementing machine learning algorithms for embedded devices is increasing more and more for utilization in real society. However, GPGPU widely used in machine learning algorithms makes it difficult to satisfy constraints on embedded devices. We have proposed Linear Array IMAX which one dimensionally arranges units with arithmetic unit and local memory that are more power efficient and area efficient while securing programming ease. However, the startup overhead during execution of multiple loops was a problem. In order to reduce the startup overhead, simultaneous execution of multiple loops and updating of intermediate results are necessary. In this research, a feedback path is provided in IMAX operation unit, improvement is made to enable multiple loop control and local memory read-modify-write, and using ARMv8 + IMAX prototype using FPGA SoC and large scale FPGA Evaluation was done. As a result, it was revealed that the improved version of IMAX on the prototype system has execution performance of 4.28 times in matrix product and 5.38 times in convolution operation, compared with conventional version IMAX. language of the presentation: Japanese 発表題目:ユニット内フィードバックによるリニアアレイの多重ループ対応手法 発表概要:近年，計算機の性能向上によりDeep Learning等機械学習アルゴリズムの性能が大きく向上している．実社会の活用のため，組み込み機器における機械学習アルゴリズム実装の需要はますます高まっている．しかし，機械学習アルゴリズムで広く利用されているGPGPUでは組み込み機器における制約を満たすことが困難である．我々は，プログラミング容易性を確保しつつ，電力効率と面積効率に長けた演算器とローカルメモリを備えるユニットを一次元に配置したリニアアレイIMAXを提案してきた．しかし，多重ループ実行時の起動オーバヘッドが課題であった．起動オーバヘッド削減のためには多重ループの一括実行と途中結果の更新が必要である．本研究では，IMAXの演算ユニット内にフィードバックパスを設け，多重ループ制御とローカルメモリのRead-Modify-Writeを可能とする改善を行い，FPGA SoCと大規模FPGAを用いたARMv8+IMAXプロトタイプを用いて評価を行った．その結果，プロトタイプシステム上の改良版IMAXで，従来版IMAXに比べて，行列積で4.28倍，畳み込み演算では5.38倍の実行性能を有することが明らかとなった． | |||
| 植田 秀樹 | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 |
|
title: Prediction of proteolytic cleavage site by γ-secretase
abstract: γ-Secretase is a membrane-embedded protease complex consisting of presenilins (PS), nicastrin, anterior pharynx defective 1 (APH-1), and presenilin enhancer 2 (PEN-2). γ-Secretase is known to cleave about 90 type-I-transmembrane proteins in their transmembrane regions but the mechanism of its cleavage site recognition is still unclear. The problem is that γ-secretase cleaves its substrates at various points, which could lead to pathological problems. For example, it degenerates amyloid beta precursor protein at various sites and produces normal amyloid beta 40 or toxic amyloid beta 42, which is believed to cause Alzheimer's disease. Therefore, it is important to predict at which sites of a substrate γ-secretase cleavage occurs. In this research I will present a cleavage site prediction model based on support vector machine (SVM) classification algorithm. It takes the amino acid sequences of the substrates already known to be cleaved by γ-secretase as the input and predict cleavage or non-cleavage as the output. language of the presentation: Japanese | |||
| 大久保 達矢 | 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: Development a supporting system on shopping for visually impaired people
abstract: When visually impaired people go to shop, finding wish items is difficult for them. Therefore, they have to get supporting from them family or helper. In our research, we will develop a supporting system on shopping for them. To develop the system, we have 3 big tasks. Those are making an environmental map that is recorded positions of shelves and walls, classification the genre of a shelf from an image, navigation to a shelf that has wish items. In this presentation, we propose the system and report how to solve these tasks. language of the presentation: Japanese | |||
| 大久保 至道 | 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: A study on providing an effect to human behavior
with feedback of Self-Awarenesss
abstract: The purpose of our research is to encourage the real-self to approach the ideal-self by using self-mirror video feedback which recognize the reality self during conversation.Therefore, we propose a new visual feedback which feed back delayed image. In this presentation, we will report the result of effects of feedback on self-awareness during the interview test task. language of the presentation: Japanese | |||