特別講演会 Special Lecture Meeting
共催：日本工業大学 In cooperation with Nippon Institute of Technology
＜案内チラシはこちら(PDF file)＞
【日時：Date】
平成２８年７月３０日（土） １２時半～１７時半
July 30th (Sat), 2016, 12:30-17:30
【場所：Place】
日本工業大学 工学部 ＬＣセンター，マルチメディア教室
Multimedia Theater Room, LC Center, Nippon Institute of Technology
Access アクセス：
（Japanese 和文） https://www.nit.ac.jp/guide/
（English 英文）http://www.nit.ac.jp/english/nit_campus/
Info. of LC center:
https://www.nit.ac.jp/center/live/lc.html
【内容：Contents】
(0) Opening [12:30-12:40]
(1) Special Lecture 1 [12:40-13:40]
Prof. Arnaud Doucet, Oxford University, UK.
Title:
On a novel class of pseudo-marginal methods for Bayesian inference
Abstract:
Pseudo-marginal methods are a very popular class of Markov chain Monte
Carlo schemes used to perform Bayesian inference for complex statistical
models. It has found thousands of applications in econometrics, epidemiology,
genetics etc. However, these iterative algorithms are very computationally
expensive, having a computational complexity scaling quadratically with
the number of data points at each iteration. I will introduce a new class
of pseudo-marginal algorithms which rely on some simple variance reduction
ideas. I will provide theory showing that these novel schemes scale better
with the number of data points. In our numerical examples, the efficiency
of computations is increased relative to standard pseudo-marginal method
by orders of magnitude for large data sets.
(2) Special Lecture 2 [13:50-14:30]
Dr. Gareth Peters, University College London, UK.
Title:
New Functional and Matrix Variate Regression Approaches for Spatial and Temporal Settings
Dynamic Cointegration Models for Commodities and Efficient Samplers
Abstract:
In this talk two new approaches to statistical regression modelling will
be discussed. The first involves developing new regression methods for
quantile function dynamics under a Symbolic Data Analysis (SDA) formalism.
The second approach is a joint temporal panel regression combining dynamic
basis function state space models with covariance regressions. New estimation
and calibration approaches will be developed. Then two applications of
these ideas will be explored, the first in quantile dynamics that link
intra-daily volatility modelling with daily volatility. The second example
will involve stress testing methodology for multiple yield curve models,
with an illustration of the Euro-region Gilts based on credit quality and
liquidity factors.
In this talk we will discuss different dynamic cointegration state space
model formulations in a Bayesian setting. We will discuss construction
of priors on the cointegration space, dynamics in the deterministic trend
and dynamics in the covariance structure of the driving innovation noise.
A novel class of collapsed Gibbs sampler will be derived to sample efficiently
from the posterior on these state space structures. Finally, illustrations
will be provided based on the Soy bean crush spread trading strategy that
trades long and short positions of soy beans, soy oil and soy meal.
(3) General Lecture 1 [14:30-15:00]
Prof. Mitsunori Mizumachi and Maya Origuchi, Kyushu Institute of Technology,
Fukuoka, Japan
Title:
Superdirective beamforming with deep neural network
Abstract:
Beamforming has been one of hot issues in acoustic signal processing, because
it can achieve signal enhancement and sound source localization. In general,
a beamformer is designed by an analytical approach and adaptive filtering.
It is, however, difficult to properly optimize the beamformers under the
complicated acoustical scene. In this talk, a flexible framework for optimizing
the beamformer is introduced based on a deep neural network. Performance
of the proposed method is confirmed by computer simulation. The proposed
beamformer could successfully achieve superdirectivity compared with conventional
beamformers.
(4) General Lecture 2 [15:00-15:30]
Prof. Shin-ya Sato, Nippon Institute of Technology, Saitama, Japan.
Title:
Identifying structural components of complex network by signal response
analysis
Abstract:
The study of networked systems, such as Web and neural systems, has attracted
a lot of attention. One of the important subjects in this study is to understand
correspondences between functions of systems and structural characteristics
of the networks derived from the systems. Understanding of network structures
is expected to lead to understanding of complex systems. In this talk,
a unique method for analyzing structural characteristics of networks is
presented, where information on network structure is obtained by analyzing
response signals of the networks. The definition and properties of response
signals are explained using artificial networks. A method for identifying
structural components of a network by analyzing its response signal is
also presented.
(5) General Lecture 3 [15:30-16:00]
Nishanth Koganti (Nara Institute of Science and Technology), Ravi Joshi (Kyushi Institute of Technology), Tomoya Tamei (NAIST), Kazushi Ikeda (NAIST), Tomohiro Shibata (KIT)
Title:
Bayesian Nonparametric Latent Manifold Learning for Robotic Clothing Assistance
Abstract:
Robotic clothing assistance is a challenging problem involving close interaction of the robot with non-rigid clothing articles and with the assisted person whose posture can vary during assistance. Design of an efficient clothing assistance framework involves reliable state estimation handling the nonlinear dynamics of clothes and data-efficient motor skills learning for reliable interaction with humans. In this talk, we present the application of Bayesian nonparametric latent manifold learning to two problems of robotic clothing assistance. In the first case, we address the problem of reliable cloth state estimation. We propose the use of Manifold Relevance Determination (MRD) to learn a shared latent manifold for observations from a noisy depth sensor and accurate motion capture system. This latent manifold is reliably used to infer the accurate cloth state given only the noisy depth sensor readings in a real-world setting. In the second case, we address the problem of efficient motor skills learning for clothing assistance. We propose the use of Bayesian Gaussian Process Latent Variable Model (BGPLVM) to learn a low dimensional latent manifold, encoding the motor skills for clothing assistance performed by a dual-arm 7 DOF robot. This latent manifold is shown to generate high dimensional clothing trajectories that not only follow task space constraints such as the coupling with clothes but also generalize to unseen environmental settings. This representation is also shown as an alternate approach to learning from demonstration to recover from various failure scenarios.
(6) General Lecture 4 [16:00-16:30]
Prof. Kazuhiko Kawamoto and Yoshiyuki Tomura (Chiba University), Japan.
Title:
Learning pedestrian dynamics with Kriging
Abstract:
Modeling pedestrian dynamics have numerous applications such as tracking, navigation, group behavior analysis, and abnormal behavior detection. This talk presents a method for learning pedestrian dynamics with Kriging. Kriging is a spatial interpolation method widely used in geosciences. Pedestrian dynamics is generally restricted by other pedestrians and its restriction is caused by social interaction between them. In the proposed method, the social interaction is represented by spatio-temporal correlation of pedestrian dynamics and the correlation is estimated by Kriging.
(7) General Lecture 5 [16:30-17:00]
Dr. Kuniyoshi Hayashi, St. Luke's International University, Tokyo, Japan.
Title:
Stability Analysis of a Change-Point Detection Method Based on Influence Functions
Abstract:
In the field of statistics, influence functions have played an important
role in evaluating the effect of an observation on target statistics or
statistical models. For such traditional statistical diagnostics, we usually
assume that the properties of population parameters of data to be analyzed
are fixed in the assessment based on influence functions. Therefore, if
the population parameters are time-dependent, we need to detect a change
point where the properties of population change. By properly shifting the
population parameters at the change point, we can exactly evaluate the
influence of observations on target statistics or statistical models to
be analyzed on the shifted population parameters. For the above situations,
we have developed methods for detecting a change point by using influence
functions. Moreover, we have investigated the performance of the proposed
method through some numerical examples. In this paper, we propose a framework
of a computational stability analysis for our change-point detection methods
based on influence functions. We confirm the effectiveness of our proposed
approach through the results of numerical examples.
【備考：Remark】
開催当日は日本工業大学のオープンキャンパスが開催されており，本行事に見学者の来訪が有り得る事をご承知おき下さい．
Open Campus event will be held on the day, so visitors to the event may
observe the lecture talks.
【意見交換会：Dinner Table for Discussion】
講演会の終了後に開催の予定（実費）．
Dinner Table for Discussion will be taken place after the seminar with
participants' own expense.
|
|
