ゼミナール発表

日時: 12月11日(月)3限(13:30-15:00)


会場: L1

司会: 小林 泰介
伊藤 健史 1751008: M, 1回目発表 数理情報学 池田 和司☆
title: A Structured Deep Neural Network Architecture to Learn Robot Dynamics
abstract: Modelling dynamics of robots is essential for precise motion control. However, some nonlinear physical phenomena such as contact or friction are still challenging to be learned with parametric models. Recently, deep neural networks (DNNs) are commonly used as non-parametric models to approximate nonlinear functions. DNNs’ capability to model complex functions is quite good, but they require large training datasets to acquire good approximation. In this study, we developed a novel structured deep neural network architecture to learn robot dynamics with fewer training samples, by using prior knowledge about mechanical structure of robots. We will evaluate its performance on a simulated environment.
language of the presentation: Japanese
 
井上 嵩史 1751009: M, 1回目発表 数理情報学 池田 和司☆
title: Population prediction by using LSTM
abstract: In these years, urbanisation is accelarated. According to the population increasing, the crowdedness and delays are caused by people's complicated moving. In my research, to solve these problems, I predict a population of a specific area. In progress of my research, I have predicted with VAR model(Vector Auto Regressive model). To improve the accuracy, I will introduce a model using LSTM(Long Short-Term Memory).
language of the presentation: Japanese
発表題目: LSTMを用いた人口予測
発表概要: 近年, 都市化が加速している. 人口の増加と共に個々人の複雑な移動によって混雑, 遅延が突発的に生じている. 本研究では, これらの問題解決のため特定の地区の人口を予測する. 研究の進捗としては, まずVARモデル(Vector Auto Regressive model)を用いて予測した. そこで, さらに予測精度向上を図るためLSTM(Long Short-Term Memory)を用いた予測手法を紹介する.
 
尾崎 隼平 1751025: M, 1回目発表 数理情報学 池田 和司
TItle: Neural circuit mechanism that produces facial expression of mouse
Abstract : Our goal is to clarifies the neural circuit mechanism for emotion to manifest as facial expression. You know that emotions appear in your face like a happy face or a sad face. However, systematic research has not been done on emotional expressions by facial expressions of animals. Also, understanding of the neural circuit mechanism that creates expression and its evolution is not advanced at all. Therefore, in this study, We artificially manipulate the activities of nerves that control the feelings of animals. Then we investigate the changes of various facial muscle movements and clarify the neural circuit mechanism for emotion to manifest as facial expression. Among them, today, we estimate the feelings of mouse from whisker movement.
Language of the presentation: Japanese
 
譽田 実希子 1751042: M, 1回目発表 数理情報学 池田 和司
Title: Analysis of Finless Porpoise Hunting Behavior
Abstract: Finless porpoise, Neophocaena asiaeorintalis, is small-toothed whale distributed in coastal water in Japan, China and Korea. They are not gregarious animal and it is rare for them to make large flocks. However, they sometimes form large group to hunt fish. Their forming strategy affecting is unknown. We try to analysis of finless porpoise hunting behavior. We extracted porpoises from their feeding movie and we construct a mathematical model that matches it.
Language of the presentation: Japanese
 
鈴木 文丈 1751057: M, 1回目発表 数理情報学 池田 和司☆
TItle:Choice-induced preference change through multi-dimensional approach
Abstract: Cognitive dissonance can be expressed as discomfort when we face a fact that contradicts our beliefs or attitudes. Choice-induced preference change (CIPC) can be examined by the “free-choice paradigm”, of cognitive dissonance theory. For example, after we choose either of two equally preferred items, we prefer the chosen item and we dislike the item not chosen. That is because we tend to justify past behavior by changing our attitude. Our preference is considered to consist of various factors. On the other hands, preference of previous studies is defined as values that was reported how much they prefer in showing the images of food and tourist spot. In this time, we define preference as multi-dimension using emotional component extracted from image database in addition to self-report. And then, we evaluate the effect of preference change. Specifically, we predict preference change from brain activity during execution of free choice paradigm, and then examine which component contribute to the prediction.
Language of the presentation: Japanese
 
高橋 晴太郎 1751061: M, 1回目発表 数理情報学 池田 和司
title: Dialogue Breakdown Detection with Conversational Model
Dialogue breakdown detection is NLP task that detect wheter a system utterance causes breakdown. If a system can detect inappropriate utterance, it can avoid it and reply better one. We use hierarchical LSTM to get vector representations of sentences with long term context.
language of the presentation: Japanese
 

会場: L2

司会: 進藤 裕之
池内 加奈 1751005: M, 1回目発表 知能コミュニケーション 中村 哲
title: Behavior decision on utterances not coming out intentions
abstract: Recently, there are many dialogue systems that can understand and respond to human utterances. But most of these systems deal only fixed answers or obvious utterances. In this research, we aim at a dialogue system that can take appropriate actions on utterances containing implicit intentions or not having goals.
language of the presentation: Japanese
発表題目: 意図が表出しない発話に対する行動の決定
発表概要: 近年、人間の発話を理解して返答できるような対話システムが数多く存在するが、ほとんどのシステムは返答が定型的であったり、 発話の表面上の意味しか理解することができない。 そこで、本研究ではユーザの暗黙的な意図が含まれた発話やユーザが意図そのものを意識していない発話に対して適切な行動がとれる 対話システムを目指す。
 
宇城 毅犧 1751019: M, 1回目発表 知能コミュニケーション 中村 哲
title: Detection of dementia from various questions
abstract: Japan became a super-aging society due to the increased elderly population. This is associates with the increasing of the number of dementia patients, which is considered a serious social issue. Detection of dementia requires through examination such as blood tests and fMRI, but this process provokes a great deal of anxiety and stress for those who having the examinations. Previous studies proposed the screening for an easy detection of dementia. These studies focus on the content of utterances to detect dementia, whereas this research uses dialogue with computer avatar in order to detect dementia.
language of the presentation: Japanese
 
勝見 久央 1751028: M, 1回目発表 知能コミュニケーション 中村 哲
title: Optimization of Evidence Collection Dialog for Generating Rational Arguments
abstract: In this work, I am trying to optimize evidence collection dialog. It is frequently observed among argumentation based dialog, such as persuasion dilalog and inquiry dialog etc, whose purpose is defeating other pariticipants with their arguments. In this kind of dialog, better evideces make the arguments more rational. Therefore it is important to optimize the policy how to collect appropriate evidence for supporting the arguments. In order to optimize it to build rational arguments efficiently, I hove formulated evidence collection dialog as Markov Decision Process, executed simulation with reinforcement learning, and evaluated the learned policy.
language of the presentation: Japanese
発表題目: 論証構築のための情報探索対話戦略の最適化
発表概要: 本研究では証拠収集対話の最適化を試みる。 証拠収集対話とは、説得対話や審理対話といった、論証のやり取りによって成立するような多くの議論対話において、 対話参与者が自らの主張の根拠とする証拠を収集するために、 他の対話参与者や第三者に問い合わせたりするような対話である。 この証拠収集対話をマルコフ決定過程として定式化し、 合理的な論証を構築するための効率的な対話戦略の最適化に取り組んでいる。また、強化学習によって得られた最適化戦略の評価を行った。
 
柴田 敦也 1751053: M, 1回目発表 知能コミュニケーション 中村 哲
title: Programming support by using code which was stored OJS.
abstract: In many cases, simple code is already written by someone. If you could use/refer existing code, you can write programs more easily. For example, programming library is one of the technique using existing code. In this research, I seek small snippet or non-named but high frequency appearing snippet from OJS(OnlineJudgeSystem). And I aim to construct a system which suggests snippets which close to user's desired processing.
language of the presentation: Japanese
発表題目: OJSの保有するコードを用いたプログラミング支援
発表概要: 簡単な処理を行うコードを書きたいと考えた時、その処理は過去に誰かの書いたことのある処理である場合も多い。もしも、その既に書かれたコードを使用/参照することができれば、プログラミングの難易度を下げることができる。例えばプログラミングライブラリも用いたプログラミングは、既存のコードを利用する方法の一つである。本研究では、一般的なライブラリよりも小規模、また名前は付いていないが頻繁に利用される処理の塊(スニペット)をOJS(オンラインジャッジシステム)の保有するコードから見つけ出し、ユーザの目的とする処理に近いスニペットを提示するシステムの構築を目指す。
 
石橋 優希 1751006: M, 1回目発表 知能コミュニケーション 中村 哲☆
Title:Decision-making change in the elderly due to cognitive impairment: functional brain and behavioural model
Abstract: In our daily life, we make decisions in various situations. How to make such decisions differs among elderly people,young people,patients with psychiatric disorders,etc. Due to differences in brain activity and binding. In this experiment,from the viewpoint of model base and model free,to observe the decrease of model base of elderly people.
Language of the presentation: Japanese
 
竹内 瞭 1751065: M, 1回目発表 知能コミュニケーション 中村 哲☆
title:Multi Liner Regression of Symptom-related Tweets for Infectious Gastroenteritis Scale Estimation
abstract:To-date various Twitter-based event detection systems have been proposed. Most of their targets, however, share common characteristics. They are seasonal or global events such as earthquakes and u pandemics. In contrast, this study targets unseasonal and local disease events. Our system investigates the frequencies of disease-related words such as “nausea,” “chill,” and “diarrhea” and estimates the number of patients using regression of these word frequencies. Experiments conducted using Japanese 47 areas from January 2017 to April 2017 revealed that the detection of small and unseasonal event is extremely difcult (overall performance: 0.13). However, we found that the event scale and the detection performance show high correlation in the specied cases (in the phase of patient increasing or decreasing). The results also suggest that when 150 and more patients appear in a high population area, we can expect that our social sensors detect this outbreak. Based on these results, we can infer that social sensors can reliably detect unseasonal and local disease events under certain conditions, just as they can for seasonal or global events.
language of the presentation:Japanese
 

会場: L3

司会: Duong Quang Thang
世良 京太 1751058: M, 1回目発表 ロボティクス 小笠原 司
title: Robust object grasping with 3D-CNN in an occluded area
abstract: In an environment where several kinds of objects are placed cumbersome, in order for the robot to stably grip a designated object, it is necessary to recognize the shape and occluded area of the object. In previous research, there are only things that grasp several kinds of simple objects, and those that recognize objects with 3D-CNN learned by image information and depth information. In this research, several objects recognized using 3D - CNN are removed by using a dual arm robot and gripping of a specific object.
language of the presentation: Japanese
 
土屋 裕杜 1751071: M, 1回目発表 ロボティクス 小笠原 司
title: Pouring from a deformable container based on haptic information by a double-arm robot
abstract: Research on robots for domestic tasks in place of us becomes popular. Many domestic operations, such as cooking and refilling detergent, involve pouring motion. So far, the two research directions were tried to achieve the robot. One direction learns the skill to pour various ingredients from human demonstration. The other direction estimates the amount of the pouring ingredients using thermal and color images. These directions consider the task to pour from a rigid container such as a stainless steel bottle or cup and do not consider from a deformable container such as a retort pouch or a plastic bag. In this study, we propose a method to pour liquid or beans from a deformable container based on haptic information by a dual-arm robot. Haptic information makes grasping the container robust against change of shape and center of gravity.
language of the presentation: Japanese

 
中村 匠 1751085: M, 1回目発表 ロボティクス 小笠原 司
title: Operating home appliances by a robot using knowledge database
abstract: In a home environment, operating home appliances is very common. To achieve the operation by a robot, machine learning-based recognition is employed, since such recognition can recognize the scene robustly. However, it is very time-consuming to train the recognizer. In this research, we propose a robot system that learns the generic and task-specific knowledge separately. By sharing the generic knowledge among the various tasks, we try to reduce the effort for learning. Constructing this system requires the system to recognize the surrounding objects. We show the current progress of implementing the system.
language of the presentation: Japanese
 
樋口 太也 1751098: M, 1回目発表 ロボティクス 小笠原 司
title: Development of 2 degree of freedom active ankle-foot orthosis for traverse walking
abstract: During the forestry work, workers should walk in the mountain not only up or down but also sideways. Such traverse walk is very hard for them due to the large burden around the ankle joints. In this research, in order to support such traverse walk, we improve our power assist suit with using active ankle-foot joints. We design the device by considering the degrees of freedom in both the sagittal plane and the frontal plane to make the walking assistance comfortable. We also control the assisting torques for each plane based on the measured joint torques of human ankles when walking without the assistance. We measure torque of human ankles by inputing motion data and external force into musculoskeletal simulator. We evaluate assist efficiency based on muscle activity.
language of the presentation: Japanese
 
佐藤 諒 1751051: M, 1回目発表 ロボティクス 小笠原 司
title: Artificial pollination by multicopter
abstract: Large farming requires efficient pollination. Though one of the solutions is pollination by bees, bee supply is unstable due to climate change and pesticide influence. Since increasing population in global needs new agricultural lands, enlargement of farming is inevitable. If we would realize a robot that can pollinate, we can replace the bees and reduce the burden on people for expanding the agricultural environment. In this research, we propose artificial pollination by a multicopter.
language of the presentation: Japanese
 
山根 弘樹 1751119: 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: Development and Evaluation of Neural Networks using Oxide Semiconductor Synapses for Letter Reproduction
abstract: Neural networks using oxide semiconductor synapses are developed and evaluated for letter reproduction. It is assumed that amorphous metal-oxide semiconductor devices are used for the synapse elements, and the characteristic degradation is utilized for the learning rule named modified Hebbian learning. First, we explain architecture and operation of a Hopfield neural network, simulate the letter recognition by the neural network, and show a degradation map. Particularly in this presentation, we explain the simulation algorithm in detail. Next, we explain architecture and operation of a cellular neural network, and simulate them. Particularly in this presentation, we explain the simulation algorithm in detail also for the cellular neural network. In addition. we compare the Hopfield and cellular neural networks, and it is found that the former has higher performance, although the latter has a simple structure.
language of the presentation: Japanese