| 山之口 智也 | M, 1回目発表 | 知能システム制御（ロボットラーニング） | 杉本 謙二☆, 池田 和司, 松原 崇充 |
|
title: Real-to-Sim Image Transfer for Vision-based Robot Control
abstract: Learning robot action in simulation is much easier than doing so in the real-world. However, applying the learned control policy in a simulator to the real robot suffers from a severe problem, so-called reality gap. In particular, the generalization of the vision-based policy is challenging. In this study, we focus on the approach of real-to-sim image transfer, which is one of the promising methods to overcome this reality gap. Then, we explore how to leverage the knowledge of robot dynamics to improve the accuracy and data efficiency of the transfer. Our preliminary results in simple simulation suggest the effectiveness of our approach.; language of the presentation: Japanese | |||
| 鹿内 裕介 | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, 加藤 博一, 酒田 信親, 磯山 直也 |
|
title: Examining the introduction of Third Person’s View experience in the real world
abstract: The third-person viewpoint is often used as the player’s operation viewpoint in video games. It can present the player with both the own character and its surroundings simultaneously. In this study, I investigate behavioral and sensory changes caused by experiencing a third-person viewpoint in a real environment. In the proposed system, a camera is attached to the tip of a stick that extends backward from the user, and the captured image is displayed on the HMD that the user is wearing. To solve the problem that the user’s front is shielded by himself, I arrange an additional camera in front of the user’s chest and display the shielded area transparently. language of the presentation: Japanese | |||
| 廣瀬 雄士 | M, 1回目発表 | 自然言語処理学 | 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之 |
|
title: Knowledge base completion with relation path information
abstract: Knowledge base completion(KBC) is a task to predict the missing part of an incomplete triples such as (Obama, livesIn, ?) using preexisting knowledge base(KB). A common approach is to embed entities into a vector space and model relations as matrices to calculate triplet scores(bilinear model). This approach is reasonable to capture the features of KB, but relation matrices have much parameters, which often causes overfitting. Another problem is how to learn the relation path features from KB. To overcome these problems, previous research proposed a model which adds a path score to the score of bilinear model. This approach can reduce relation parameters and consider paths around triplets. In this study, I propose to combine reinforcement learning with the approach. language of the presentation:Japanese | |||
| 江口 竣 | 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, 安本 慶一, 神原 誠之, Alexander Plopski, 藤本雄一郎 |
|
title: Diet Therapy Support System for CKD Patients
abstract: Chronic kidney disease is one kind of common disease. The number of CKD patients is equal to one-eighth of Japanese adult and it is said CKD can’t be recovered. Diet Therapy has important role for CKD treatment. However, it is difficult to implement diet therapy in long-term, because every day meal management is required and weighing takes long time. In this research, we propose a support system that can help in meal management through weighing each specific ingredients. Then, CKD patients can implement diet therapy in long-term by using this. language of the presentation: Japanese 発表題目: *** この部分を発表題目に *** 発表概要: *** この部分を発表概要に *** | |||
| 千田 将也 | M, 1回目発表 | ソフトウェア設計学（超高信頼ソフトウェアシステム検証学） | 飯田 元, 松本 健一, 片平 真史（客員）, 吉石濱 直樹（客員），高井 利憲（客員） |
|
title: Towards eXplainable AI taking user's requirements into account
abstract: In recent years, eXplainable AI has attracted attention. With this technology, it is possible to explain the prediction of black box machine learning algorithms, and expected to be used for supporing decision-making by human beings and validating machine learning models in system development phase. Currently, many explainable AI methods have been proposed in academia. On the other hand, for practical use the users generally have some purpose and requirements on explanation but the selection of suitable explainable AI methods for the user's requirements has not been well-studied. In this work, we first classify user's explanation requirements based on the framework of perceptual uncertainty of systems using machine learning technologies and then investigate suitable explainable AI methods for each requirement class. language of the presentation: Japanese | |||
| SALDAJENO DON PIETRO | M, 1回目発表 | 数理情報学 | 池田 和司, 金谷 重彦, 吉本 潤一郎, 久保 孝富, 福嶋 誠 |
|
title: Mathematical Modeling of the Circadian Clock: Potential Applications in Cancer Therapy
abstract: The circadian clock is the natural biological clock that controls the sleep-wake cycle and many other functions in the human body. This biological clock is driven by the interaction between a group of genes known as the "core clock genes". Aside from interacting with each other, these core clock genes also control many other downstream target genes, some of which have been implicated in cancer cell growth and cancer metastasis. Therefore, the genes of the circadian clock are potential targets for cancer therapy, to suppress the growth of cancer cells and to help prevent metastasis. To determine the most efficient and effective ways to control the downstream target genes (and thereby control cancer growth and metastasis), it is necessary to use mathematical modeling.This research project aims to expand existing mathematical models of the circadian clock to include the relationship between the core clock genes and the downstream target genes that are related to cancer cell growth and metastasis. Gene expression data from biological experiments will be used as the basis for expanding the currently existing mathematical models. language of the presentation: English | |||
| 野村 武司 | 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, 林 優一, 中田 尚, Tran Thi Hong, 張 任遠 |
|
title: Spiking Neural Networks with Efficient Re-configurability
abstract: Spiking neural networks (SNNs) which closely emulate the behaviors of natural neural networks are investigated in this work. Representing the data in terms of the spike-coding, SNNs appear the potential for performing variant functions as the real brain. On the other hand, their specific hardware implementations are demanded instead of numerically implementing them only via software since the models are always described by differential equations. To extend the specific hardware implementations of SNNs into more general purpose application fields, the post-silicon reconfigurable SNN processors are expected. Unfortunately, the conventional topology of neural networks known as full-connection (FC) leads to remarkable redundancy on synopses and neurons during re-configuration, which are expensive on hardware resources. In this study, an innovative topology of neural networks is developed in the fashion of bisection connection. By migrating the SNNs into the proposed bisection neural network, the entire SNN is expected to post-silicon reconfigure efficiently. Both of mathematical model and circuit implementation are investigated for the novel SNN topology. language of the presentation: Japanese | |||
| 砂田 翼 | M, 1回目発表 | ソフトウェア工学 | 松本 健一, 飯田 元, 石尾 隆, 畑 秀明, Raula G. Kula |
|
title: Towards cheating detection in programming tests
abstract: In company recruitment activities, an increasing number of companies use programming tests to evaluate the technical capabilities of candidates. However, cheating is performed in the test, and there is a problem that the candidate's technical ability cannot be evaluated correctly. The purpose of this study is to detect cheating behavior in coding tests. High-speed and high-accuracy are required for cheating detection. Therefore, we propose a method for detecting the similarity of source code by using the source code obtained in the actual company coding test and detecting cheat from the similar source code. language of the presentation: Japanese | |||
| 橋本 敬志 | 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, 久保 尋之, 田中 賢一郎 |
|
title:Epipolar imaging under high-frequency illumination for highlighting interior of translucent objects
abstract:Estimation of structures in a translucent object by imaging is required in various tasks, for example, foreign substances inspection and visualization of human inside. Since Episcan separates indirect reflected light from scene using epipolar geometry, we can capture indirect component only and get images in which the interior is highlighted. However, the images are still unclear because captured images contain scattered light component. It is required to capture clearer images for more precise estimation of the interior. In this study, we propose a sharp imaging method that removes scattered light by computation of images projected checkerboard pattern as light source. language of the presentation: Japanese | |||
| 岸本 泰海 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, 加藤 博一, 舩冨 卓哉, 久保 尋之, 田中 賢一郎 |
|
title:Appearance Analysis of Metals using Microgeometry for Aesthetic Evaluation
abstract:Metal parts inspection contain evaluation of shape, scratches and appearance. Above all, aesthetic evaluation is important for appearance inspection in order to determine whether unintended scratches can be allowed. We can detect scratches on metal surfaces using image processing technology, but it is difficult to predict the appearance of the scratches. In recent years, Dong’s group proposed a method for calculating high-precision normal distribution function (NDF) from microgeometry. In this study, I look for a method to predict the material that matches the appearance from the measured high-precision NDF. In this presentation, I show the result of calculating NDF from the height map for predicting appearance. As a result of experiment, the difference of NDF by processing process was confirmed. language of the presentation:Japanese 発表題目:審美的評価のためのマイクロジオメトリーを用いた金属の外観分析 発表概要:金属部品の検査では、形状や傷・外観の評価がある。中でも外観検査では、意図しない傷を許容するかどうかの判定には、審美的な評価が重要である。画像処理技術を用いることで金属表面の傷を検出するができるが、傷の見た目を予測することは難しい。近年、Dongらのグループはマイクロスケールのジオメトリをもとに高精度の法線分布関数(NDF)を計算する手法を提案した。本研究では、この手法によって計算された高精度のNDFから見た目と一致する質感を予測する手法を模索している。本発表では、金属の外観予測に向けて、高さマップからNDFを計算した結果を示す。実験の結果、加工プロセスによるNDFの違いを確認した。 | |||
| 本田 卓 | 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, 井上 美智子, 中田 尚, Tran Thi Hong, 張 任遠 |
|
title:Implementation of Variational Bayesian Gaussian Mixture Models on IMAX
abstract: Gaussian Mixture Models (GMM) are a representation method that realizes clustering, and widely used in many applications for probability density modeling and soft clustering. However, the parameters of GMM require estimating from data. Variational Bayesian Gaussian Mixture Models (VBGMM) are parameter estimation methods that can avoid over-fitting than other methods. We have developed In-Memory Accelerator eXtension (IMAX). IMAX is an accelerator that contains Coarse Grain Reconfigurable Arrays (CGRA) architecture and each execution units contain local memory. In this presentation, we show the outline of implementation of VBGMM on IMAX. language of the presentation: Japanese | |||