| NGUYEN LE AN | 1661029: D, 中間発表 | 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之, 能地 宏 |
|
title: Improving Sequence to Sequence Neural Machine Translation by Utilizing Syntactic Dependency Information
abtract: Sequence to Sequence Neural Machine Translation has achieved significant performance in recent years. Yet, there are some existing issues that Neural Machine Translation still does not solve completely. Two of them are translation for long sentences and the over-translation. To address these two problems, we propose an approach that utilize more grammatical information such as syntactic dependencies, so that the output can be generated based on more abundant information. In our approach, syntactic dependencies is employed in decoding. In addition, the output of the model is presented not as a simple sequence of tokens but as a linearized dependency tree construction. Experiments on the Europarl-v7 dataset of French-to-English translation demonstrate that our proposed method improves BLEU scores by 1.57 and 2.40 on datasets consisting of sentences with up to 50 and 80 tokens, respectively. Furthermore, the proposed method also solved the two existing problems, ineffective translation for long sentences and over-translation in Neural Machine Translation. language of the presentation: English | ||
| 佐藤 元紀 | 1651053: M, 2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, 能地 宏 |
|
title: Cross-Domain Parsing with Adversarial Training
abstract: Syntactic parsers are essential for many natural language processing applications, but training them requires expensive annotations of syntactic trees by linguistic experts. To ease the annotation effort, we focus in this paper on supervised paresr domain adaptation, in which only a limited amount of annotated trees on the target domain is available. Many recent state-of-the-art parsers are neural parsers and utilize word representations conditioned on the entire sentence obtained with bi-directional LSTMs. We pursue a better domain adaptation technique for such modern neural parsers. language of the presentation: Japanese | ||
| 澤山 熱気 | 1651056: M, 2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, 能地 宏 |
|
title: Improving Named Entity Recognition on Non-immediate Response Setting
Named entity recognition (NER) plays a central role of automatic knowledge extraction from scientific and technical papers. Generally, we develop a NER module independently from the overall system. At that time, we assume that the system is a ``real-time'' system, and thus, the NER module needs to respond inputs immediately as possible. However, in the scenario of the automatic knowledge extraction, the immediate response is unnecessary in most cases since the purpose of the task is to (gradually) accumulates knowledge from the papers over a relatively long time interval. From this background, we discusses approaches suitable for tackling such non-immediate response setting of NER. We specifically focus on the methods that utilize a given test data for further improving the task performance. We also reports the current progress, and preliminary experimental results on simple baseline methods and our proposed method. language of the presentation: Japanese | ||
| 寺西 裕紀 | 1651075: M, 2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, 能地 宏 |
|
title: Coordinate Structure Analysis using Dependency Tree
abstract: We propose a neural network model for coordination boundary detection. Our method relies on two common properties - similarity and replaceability in conjuncts - in order to detect both similar and dissimilar pairs of conjuncts. The model improves the identification of clause-level coordination using bidirectional recurrent neural networks incorporating two properties as features. We show that our model outperforms existing state-of-the-art methods for the coordination annotated Penn Treebank and Genia corpus without any syntactic information from parsers. language of the presentation: Japanese | ||
| 今林 亘 | 1651017: M, 2回目発表 | 知能システム制御 杉本 謙二, 岡田 実, 松原 崇充, 小蔵 正輝, 小林 泰介 |
|
title: Tolerance to Tempolal Sesing Failure in Feedforward Learning Control
abstract: This paper proposes a control system design against temporal sensing failure. Such failure is caused by occlusion by obstacles in non-contact type sensor, signal transmission loss in wireless communication or other channels by shared resources, etc. Two degrees of freedom (2DOF) structure makes it possible todesign feedback control that is robust to sensing failure and feedforward control that provides a good tracking property. The former is attained by means of a state observer whose stability is guaranteed against sensing failure under a certain mild assumption. The latter is implemented with a parameter tuning technique that the authors' group developed, which has been motivated by a biological model for voluntary motion. Numerical simulation is carried out to verify the effectiveness of the proposed control system. language of the presentation:Japanese 発表題目: 一時的なセンシング障害に耐性のあるフィードフォワード学習制御 発表概要: 非接触型センサを用いたセンシングは，センサが自由に設置でき，広範囲を計測できる利点から，自動運転やロボット制御への応用が期待できる．しかし，欠点として，センサと計測対象物の間に障害物が通過することで信号が一時的に途絶えることがある．一時的な障害なので，直ちに復旧するとしても，リアルタイム性を要求される制御においては致命的な問題となる．このセンシング障害に対して，本研究では， 2 自由度構成を用いた対策を提案する．フィードバック制御器として，一時的なセンサ障害においても安定性が確保できるような設計法を援用する．具体的には，センシング信号をサンプリングし，状態を推定して，障害が発生したときには，直前の信号を再利用して状態推定を継続する．その結果，サンプリング周期が不規則に変化するので，その場合でも誤差系が安定となるような，オブザーバゲインを設計し，状態推定値をフィードバックして閉ループ系を安定化する．またフィードフォワード制御器として，生体の運動学習から着想を得た，フィードバック誤差学習による制御器のオンライン係数調整を援用する．これにより，応答特性の改善が可能となる．これらの手法を 2 自由度構成として統合した提案法に対して，数値シミュレーションで有効性を検証する． | ||
| 友近 圭汰 | 1651077: M, 2回目発表 | ロボティクス 小笠原 司, 佐藤 嘉伸, 高松 淳, 丁 明 |
|
title: Musculoskeletal Model of Human Hand Based on its Joint Structure
bstract: Numerous robot hands have been developed so far. However, there is not yet anything that can be able to realize all of human dexterity. In order to develop a dexterous robot hand, it is necessary to understand the movements of the hand at the muscle level. Therefore, in this research, we aim to construct a musculoskeletal model of the thumb based on the joint structure in order to perform more accurate muscle motion analysis. In this presentation, we report the proposal and the progress of the joint axis estimation method necessary for constructing the musculoskeletal model. language of the presentation: Japanese | ||
| 吉岡 大輝 | 1651120: M, 2回目発表 | ロボティクス 小笠原 司, 佐藤 嘉伸, 高松 淳, 丁 明 |
|
title:Scooping motion generation with spoon using robot arm
abstract:In recent years, the number of studies in which robots perform tasks in households is increasing. Among them, cooking robot and meal support robot need scooping motion to perform these tasks. In these robots, the scooping motion is realized using the teaching or using a specific container. However, when changing the container to be scooped, it is difficult to perform an appropriate scooping operation. Therefore, in this study, we propose an automatic generation method of scooping motion using container shape and contact force information. Specifically, using the sensing information of the RGB-D sensor and force sensor, we aim to generate a scooping motion that can correspond to multiple containers. This presentation will explain the proposed method and the progress of this research. language of the presentation:Japanese | ||
| 天満 勇介 | 1651076: M, 2回目発表 | 生体医用画像 佐藤 嘉伸, 加藤 博一, 大竹 義人, 横田 太 |
|
title: Kinematic analysis of forearm rotation by 2D-3D registration using biplane fluoroscopy
abstract: In orthopedics, the kinematics analysis of bones for quantification is important to improve the treatment method. In case of forearm kinematic analysis, several studies evaluate the rotation axis using CT and X-ray images. However, in previous studies, analysis and verification of the forearm kinematics using biplane X-ray fluoroscopy have not been reported. In this study, we propose a 2D-3D registration method using a calibrated biplanar dynamic fluoroscopy and a CT image. In this presentation, we report the results of accuracy verification using a bone phantom and the results of in-vivo 3D kinematics and rotation axis estimated from the clinical fluoroscopy images. language of the presentation: Japanese 発表題目: 2方向X線透視動画像を用いた2D-3Dレジストレーションによる前腕回旋動態解析 発表概要:整形外科領域においては治療方法の改善のために，運動の定量化を目的とした人体の動態を解析する研究が行なわれている．特に前腕骨の3次元動態解析の研究としてCTとX線動画像を用いて回旋軸を評価した研究が報告されている．しかし，従来までの研究では2方向でのX線透視動画像を用いた前腕の動態に関する解析と精度検証は行なわれていない．本研究では二方向X線装置を用いて2D-3Dレジストレーションにより前腕骨の動態解析を試みる。今回の発表では2D3Dレジストレーションの精度検証のために用いた模型骨の精度検証結果と透視動画像から推定した3次元動態と回旋軸の結果を報告する． | ||
| 鈴木 啓大 | 1651063: M, 2回目発表 | 数理情報学 池田 和司 ☆, 佐藤 嘉伸, 川人 光男(客員）, 森本 淳(客員） |
|
title: MEG Current Source Estimation Method using Brain Structural Connectivity
abstruct : MEG current source estimation problem is defined as an ill-posed inverse problem because the signal source dimension is generally presumed to be higher than the observation dimension. In order to solve this problem, many kinds of constraints have been proposed. In the classical minimum norm estimation, the estimation at each time step is independent and only the spatial spread of the source is constrained. Therefore, this estimation is not suitable for dynamics of the sources. Considering the actual dynamics, each source is connected through structural connectivity composed of nerve bundles and communicated with time delay according to length of connectivity. Therefore, our group developed the source estimation method which models the dynamics and use it as a constraint. However, it is inapplicable to complicated brain activity such as resting-state dynamics. Therefore, in this research, we propose a source estimation method which achieve robustness and high accuracy using structural connectivity as a constraint of minimum norm estimation. We show the performance of estimation using simulation data. language of the presentation: Japanese | ||
| 米川 柾 | 1651126: M, 2回目発表 | 数理情報学 池田 和司 ☆, 金谷 重彦, 川人 光男(客員）, 森本 淳(客員） |
|
title: Functional differentiation of striatum involved in retaining and revaluation of learned value
abstract: Value learning is mediated by goal-directed and habit systems in brain. Goal-directed system is also called flexible behavior which has more mental load and the ability to revaluate value learned once. By contrast, habit systems is called habitual behavior which has quick response for stimulus but can not revaluate when value changes. According to previous study focusing on primate’s brain, these systems are implemented in basal ganglia, especially striatum, separately. Specifically flexible behavior and stable one are represented in rostral striatum and caudal striatum respectively. On the basis of that, we hypothesized that rostral striatum is activated during flexible value learning task and caudal striatum is activated during stable value learning task and we try to prove our hypothesis by decoding analysis with fMRI data. language of the presentation: Japanese | ||
| 沼田 椋太郎 | 1651084: M, 2回目発表 | 数理情報学 池田 和司 ☆, 松本 健一, 川人 光男(客員）, 森本 淳(客員） |
|
title: Towards the development of a single-trial brain robot rehabilitation abstract: Neurorehabilitation is a new rehabilitation method for the people suffering from disabilities in motor function due to stroke, paralysis or etc. This method applied a techniques of Brain Computer Interface (BCI) which connect brain and external devices to rehabilitation. By detecting the motor intention from the patient's electroencephalogram (EEG) and then giving sensory feedback to them as proprioceptive input, a brain plasticity will be induced and restore the motor function of patients. In this talk, we’ll talk about that a model for detecting patient's motor intention for neurorehabilitaion with single - trial EEG. language of the presentation: Japanese | ||
| 山本 諭 | 1651119: M, 2回目発表 | 数理情報学 池田 和司 ☆, 松本 健一, 川人 光男(客員）, 森本 淳(客員） |
|
title: Decoding individual's time-preference from brain activity pattern under the effect of emotional primes
abstract: We make decisions based on rewards available. At that time, We always consider the time to get the reward. It is known as an inter-temporal decision-making problem. At that problem, the temporal discounting rate which indicates the extent of how long people can wait rewards available in the future is used as an trait-like characteristic. In addition, this temporal discounting rate has not only trait-like characteristics but also state-like characteristics. In the previous study, inter-temporal decision making task with positive priming were conducted and it suggested an increase in the temporal discounting rate. In addition, the fMRI experiment of neural activity in the time preference task by happiness-fear emotional priming suggested that the anterior cingulate cortex and so on are affected by fear emotional priming at decision making. However, it is still unknown whether the brain activity at that time expresses the inter-temporal action selection modulated by emotional priming, and the brain area susceptible to the priming at that time and the prediction of the action selection from brain activity has not been conducted. Therefore, we aims to clarify the neural basis of inter-temporal action selection by emotional priming. We conduct behavioral experiments and fMRI experiments of inter-temporal decision making tasks under the effect of emotional priming for healthy subjects. We estimate the temporal discounting rate as trait-like characteristic in behavioral experiment and investigate the effect of emotional priming to temporal discounting rate. In the fMRI experiment, we try to decode the action selection. Evaluation of emotional valence and arousal by questionair is also carried out at the same time. language of the presentation: Japanese | ||