ゼミナール発表

日時: 9月27日(木)3限 (13:30-15:00)


会場: L1

司会: 松原 崇充
植田隼 1151018: M, 2回目発表 杉本謙二, 湊小太郎, 平田健太郎
title: A human’s behavior estimation method using Interval Linear Dynamical System
abstract:Due to the development of the sensor networks, human behavior estimation is actively investigated from the context of pattern recognition and machine learning. Though stochastic time-series models are usually employed, the dynamics of human motion involved in a behavior may possess a deterministic nature. Therefore, we introduce a hybrid model called as Interval Linear Dynamical System for the modeling of human behavior. It is quite similar to the motion of Jump Markov Linear Systems, known in the control community. Some preliminary results are shown in this presentation.
language of the presentation: Japanese
発表題目: 区間線形動的システムを用いた行動推定
発表概要: 近年, センサの発展に伴い, そこから得られる時系列信号を用いて人物の行動を推定し,高齢者の異常事態の検出などのサービスを提供することが注目されている. この時系列のパターン認識手法について, 信号処理的な時系列解析を用いる手法と離散状態モデルの状態遷移を用いる手法がある. 前者は複雑な動的イベントを区間に基づいて表現できず, 後者は時系列を詳細に記述できないといったデメリットがある. これらの短所を補うためにこれら二つを組み合わせたモデルとして区間線形動的システムがある. 本発表では, 人が移動する状況を対象にこのシステムを適用した結果を紹介する.
 
西田圭佑 1151078: M, 2回目発表 杉本謙二, 湊小太郎, 平田健太郎
title: Studies on Mathematical Modeling of Pressure Rice Cooking Process
abstract: We investigate the issue of mathematical modeling of pressure rice cooking. The model is useful for model-based control and model-based development of a rice cooker. To improve the model accuracy, focus on the pressurized boiling phenomenon. In this presentation, I will talk about the points for the improvement over the existing method and simulation results of boiling experiment with water only.
language of the presentation: Japanese
発表題目: 加圧炊飯プロセスの数理モデル化に関する研究
発表概要: 我々は,炊飯器へのモデルベース制御やモデルベース開発の適用を最終目標として,加圧炊飯プロセスの数理モデル化について検討している.先行研究では,米の吸水・各部の温度変化・釜内部圧力の変化を表現するモデルが提案されており,シミュレーションによって温度のみの推定精度の検証が行われている.しかしながら,釜内部圧力の推定精度については実測値との比較・検討はされていない.そこで本研究では,圧力も含めたモデルのさらなる精度向上を目的として,米を含まない水のみの沸騰実験を行い,沸騰現象に着目したモデルの改良を行う.本発表では,モデルの改良点とシミュレーション結果,今後の課題について述べる.
 
村瀬英俊 1151105: M, 2回目発表 湊小太郎, 小笠原司, 杉浦忠男, 佐藤哲大

title: Postacquisition synchronization of nongated image sequences in zebrafish myocardial cells

abstract: Zebrafish is suitable for genetics experiments, and its myocardial cells regenerate if they are damaged less than 20%. There is a demand for investigating its regenerating process especially in biological field.

Since the confocal laser scanning microscopy(CLSM) cannot scan a slice instantaneously, three-dimentional cardiac images acquired by CLSM include various cardiac phases. Conseqently, it is difficult to observe the regenaration process of the myocardial cells at the same heart phase and position.

Therefore, we synchronize nongated image sequences acquired by the CLSM,

and reconstruct the cardiac images of the same heart phase. We show the progress and the future plan of the research in the presentation.


language of the presentation: Japanese

発表題目: ゼブラフィッシュ心の非同期連続顕微鏡像に対する事後同期処理

発表概要: ゼブラフィッシュは遺伝的実験に適し,心筋細胞の20%程度を損傷しても再生する特徴を持つ.この心筋細胞が再生する過程を観察したいという要望がバイオサイエンス分野にある.

しかし,拍動の影響により,共焦点顕微鏡による一平面の撮像が瞬時に行われないことから得られる心臓の3次元画像は心拍の様々な位相が混在した画像となってしまい,同一位相・同一箇所における心筋細胞の再生過程の観察は困難である.

そこで,本研究は共焦点顕微鏡を用いて得られた非同期連続顕微鏡像に対して,事後に同期処理を施し,同位相の心臓の画像を再構成する.それによって心筋細胞の再生過程の観察を可能にするのが目的である.本発表では研究進捗と今後の課題について報告する.

 
Dy, Mary-Clare Clarin 1161204: D, 中間発表 湊小太郎, 小笠原司, 杉浦忠男
Title: 3D Optical manipulation and cell-force measurement
Abstract: Nanomanipulation offers users positional control of small objects with nanometer or submicron precision. One technique that is utilized for nanomanipulation is the optical tweezers; it can trap and manipulate small objects using a highly focused laser beam. Optical tweezers is also extensively used for the measurement of the mechanical forces of cells, wherein particles are used as probes to study and investigate cell elasticity. However, previous works typically study particle manipulation and force measurement in the lateral direction, hence excluding insights about the axial mechanical properties of cells. In our study, we have designed a scheme that can axially manipulate particles and measure the corresponding particle position with nanometer sensitivity. We translated the particle in the axial direction by adjusting the position of one lens in our setup, and then used image processing techniques to determine the axial particle movement. We investigated the capability of our system, and results show that for a given L1 lens default position, a 2.10um particle and a 4.26um particle were moved axially for 7.64 um, and 4.83 um, respectively. We also measured the axial trapping stiffness using Equipartition theorem. For a 4.26um particle, we obtained a stiffness value of 4.45uN/m for 36.75mW-laser power. After we determine the appropriate range of axial particle movement and corresponding trapping stiffness, we will compute the reactive forces of cells. Finally, this scheme will be incorporated to the existing Cell Palpation system to create a simple, yet sensitive, three-dimensional optical manipulation and cell force measurement system.
Language of the presentation: English
 

会場: L2

司会: Graham Neubig
岩尾悠祐 1151015: M, 2回目発表 鹿野清宏, 中村哲, 猿渡洋
title: Stereo Music Signal Separation Combining Directional Clustering and Nonnegative Matrix Factorization.
abstract: We address stereo music signal separation with a new method combining directional clustering and nonnegative matrix factorization(NMF). Directional clustering quantizes spatial information in the time-frequence domain by k-means like clustering, and can estimate sources and these oriental direction in multichannel signal. NMF can extract the target signal to some extent especially in the case of small number of instruments. However, for the case of many source mixtures like more realistic musical tunes, the extraction performance markedly degrades. To solve this problem, we propose a new method for musical instrument separation combining directional clustering and NMF. In the evaluation experiment, we compare directional clustering, NMF, and proposed method, and reveal that the proposed method can extract the target source with the highest accuracy.
language of the presentation: Japanese
 
大沼侑司 1151025: M, 2回目発表 鹿野清宏, 中村哲, 猿渡洋
title: High precision source separation of multiple and diffuse noise in multi-modal environment
abstract: In this study, we propose a permutation solver for performing source separation of multiple speakers, including diffuse noise with a multi-modal sensors. Blind Spatial Subtraction Array (BSSA) performs noise suppression using a noise estimator by Independent Component Analysis (ICA). However, ICA occurs permutation problem, there is a need for a permutation solver. First, we introduce how to solve the permutation problem when there are multiple speakers and diffuse noise. Next, we introduce permutation solved in the case of using a multi-modal sensors. In the evaluation experiment, we reveal that the proposed method can extract the target source with the high accuracy.
language of the presentation: Japanese
 
金原涼美 1151040: M, 2回目発表 鹿野清宏, 中村哲, 猿渡洋
title: Theoretical Analysis of Musical Noise Generation in Noise Reduction Methods with Decision-directed a Priori SNR Estimator
abstract: We propose a new theoretical analysis of the amount of musical noise generated in several noise reduction methods with a decision-directed a priori SNR estimator using higher-order statistics. In our previous study, a musical noise assessment based on kurtosis has been successfully applied to spectral subtraction. However, this approach cannot be applied to some high-quality noise reduction methods, e.g., Ephraim-Malah’s minimum mean-square error short-time spectral amplitude (MMSE STSA) estimator, because such methods include the decision-directed a priori SNR estimator, which corresponds to a nonlinear recursive (infinite) process for noise power spectral sequences. Therefore, in this paper, we introduce a computationally efficient higher-order-moment calculation method based on generalized Gauss-Laguerre quadrature. We also mathematically clarify the justification of using a typical decision-directed parameter, namely, magic number 0.98, in the decision-directed a priori SNR estimator from a viewpoint of amounts of musical noise and speech distortion. In addition, we perform comparison between these noise reduction methods based on the mathematical analysis and human perception test.
language of the presentation: Japanese
 
APRILYANTI FINE DWINITA 1151127: M, 2回目発表 鹿野清宏, 中村哲, 猿渡洋
title: Optimization Scheme of Joint Noise Suppression and Dereverberation Based on Higher-Order Statistics
abstract: Recently, a method to suppress both diffuse background noise and late reverberation part of speech has been proposed combining blind signal extraction and Wiener filtering. However, this method requires a good strategy for choosing the set of its parameters in order to achieve the optimum result. We apply the higher-order statistics parameter to automatically improve the performance of blind speech enhancement. We present an optimization scheme to control the value of Wiener filter coefficients used in this method, which depends on the amount of musical noise generated, measured by higher-order statistics. The noise reduction rate and cepstral distortion are evaluated to confirm the effectiveness of this scheme.
language of the presentation: English
 

会場: L3

司会: 竹村 憲太郎
金城健 1161005: D, 中間発表 池田和司, 小笠原司, 銅谷賢治, 吉本 潤一郎
title: Mobile robot control based on linearly solvable Markov decision process with dynamic model learning
abstract: Recently, a class of stochastic optimal control problem named as Linearly solvable Markov Decision Process (LMDP) is proposed by Todorov and his colleagues, in which the nonlinear Bellman equation is strictly linearized under some assumptions on an environmental dynamics and a cost function. Using this framework, an optimal value function and corresponding control law are obtained by solving an eigenfunction problem in continuous state and action spaces. Although LMDP framework is attractive for application to real systems like robot control, it is assumed that the dynamics of the environment is known to the robot. In addition, they just evaluated using small and simple simulated benchmarks such as pole-balancing and mountain-car tasks. To overcome this problem, we introduce a model learning component into the LMDP framework and test it in the real robot task, in which the robot learns the environmental dynamics as well as optimal control law using the sequence of experienced states and actions. We apply this method to a battery-catching task of our mobile robot named Spring Dog in which Spring Dog corrects environmental information by USB camera in a visual guide task. At first, we compare the LMDP method with the Linear Quadratic Regulator (LQR) when robot learns a model represented by a continuous-time linear model with a quadratic cost function and confirm that they obtain the same performance. In the second experiment, we prepare several basis functions to deal with nonlinear dynamics and cost function. Experimental results show that the LMDP with model learning is promising approach for robot learning in the real environment.
language of the presentation: Japanese
 
(10分待ち) __ 教員の部屋間移動のため 10分待ち
 
白石諒 1151055: M, 2回目発表 池田和司, 杉本謙二, 銅谷賢治, 吉本 潤一郎
title: Neural activity in mice parietal cortex during classical conditioning using sound source position depended reward
abstract: The parietal cortex has an important role in information processing relating to egocentric representation of target position. However, the neural circuit dynamics underlying the information processing are not well understood. And the target localization ability of mice is also not well understood. Here we conducted the classical conditioning in mice for investigating the sound source localization ability. In the experiment, the sucrose solution rewards depend on the direction and distance of sound source. We confirmed significant difference between the anticipatory licking behavior in rewarded trials and unrewarded trials. From this result, we find that mice would distinguish the different distance and direction of sound source position. Currently we are preparing optical imaging of the neural activity in parietal cortex during the task using 2 photon microscopy.
language of the presentation: Japanese
 
間島慶 1161012: D, 中間発表 池田和司, 杉本謙二, 川人光男, 神谷之康
title:Neural decoding of visual object categories using spatio-temporal patterns of electrocorticographic signals
abstract:How visual information is represented in the brain is one of the important questions in neuroscience. Some theories have suggested that brain activity patterns with high temporal resolution are used to represent visual information, which is against conventional ways base on firing rates. Although there is a lot of evidence for such temporal coding in the retina and primary visual cortex, whether visual object information in the human inferotemporal cortex depends on it or not is unknown. To answer this question, we used electrocorticography (ECoG) to measure brain activity from multiple sites with high temporal resolution and examined whether visual object categories were better predicted using fine temporal patterns of the ECoG signals. In this study, we recorded ECoG signals from five subjects. We prepared two types of input feature to statistical classifiers and compared the predictive accuracies. One type of feature is the temporal correlation between ECoG electrodes, which is an index to measure the degree of synchronization. The other is the spectral power of a signal, which ignores any phase information and fine temporal information of the signal. We found that the predictive performance using the temporal correlations better than that using the frequency powers. These results suggest that fine temporal patterns of brain activity in the inferotemporal cortex should carry information on visual object categories.
language of the presentation: Japanese