STM2015
& CSM2015
Technical Program
Location: Conference Room I (2nd floor), ISM
July 13 (Monday)
|
Time |
10:15 – 10:30am |
|
|
Opening |
|
Time |
10:30 – 12:00pm |
|
Speaker |
Prof. Makoto Maejima (Keio U) |
|
Title |
Stable distributions, stable processes
and Levy processes |
|
Abstract |
In this tutorial talk, I will explain
basic facts on stable distributions, stable processes
and Levy processes. Most parts of the talk are from
the following books: [1] G. Samorodnitsky and M.S. Taqqu,
Stable Non-Gaussian Random Processes: Sto-chastic
Models with In [2] A. Janicki and A. Weron, Simulation
and Chaotic Behavior of [3] V.V. Uchaikin and V.M. Zolotarev,
Chance and Stability: Stable Distributions and their
Applications, 1999. [4] J. Nolan, Bibliography on stable
distributions, processes and related topics, Original
version August 3, 2003, Revised June 22,
2015,http://academic2.american.edu/jpnolan/stable/StableBibliography.pdf [5] K. Sato, Levy Processes and In |
|
Time |
13:30 – 14:30pm |
|
Speaker |
Prof. Samuel Kou (Harvard U) |
|
Title |
Big data, Google and disease detection:
the statistical story |
|
Abstract |
Big data collected from the internet have
generated significant interest in not only the
academic community but also industry and government
agencies. They bring great potential in tracking and
predicting massive social activities. We focus on
tracking disease epidemics in this talk. We will
discuss the applications, in particular Google Flu
Trends, some of the fallacy and the statistical
implications. We will propose a new model that
utilizes publicly available online data to estimate
disease epidemics. Our model outperforms all previous
real-time tracking models for influenza epidemics at
the national level of the US. We will also draw some
lessons for big data applications |
|
Time |
14:30 – 15:30pm |
|
Speaker |
Prof. Frederick Kin Hing Phoa (Institute
of Statistical Science) |
|
Title |
A Swarm Intelligence Based (SIB) Method
and Its Applications to |
|
Time |
16:00 – 17:30pm |
|
Speaker |
Dr. Malcom Egan (Agent Technology Center) |
|
Title |
Symmetric Alpha-Stable Random Variables,
Fractional Calculus, and the Fourier-Mellin Triangle |
|
Abstract |
The characteristic function (the Fourier
transform of the density function) plays a key role in
the theory of alpha-stable distributions. An
alternative perspective, first investigated by
Zolotarev for stable distributions in a 1957 paper, is
based on the Mellin transform of the
density function. This Mellin-transform
perspective has intimate ties to the fractional
calculus developed by Riemann, Liouville, Marchaud and
Riesz. In this talk, we focus on the connections
via fractional calculus between symmetric alpha-stable
distributions, their characteristic functions, and
Mellin transforms of their densities. We review
and make explicit the recent use of this combination
of perspectives by Di Paolo et al to develop integral
density representations useful for numerical
approximation. We also show additional important
properties using these techniques, including
expressions for fractional moments
and probability metrics. |
July 14 (Tuesday)
|
Time |
9:00 – 10:30pm |
|
Speaker |
Prof. Nourddine Azzaoui (Université
Blaise Pascal) |
|
Title |
Identification and calibration problems
in some non-stationary alpha-stables processes |
|
Abstract |
Thanks to the spectacular progress of
modern computing, spectral analysis of stochastic
processes knew an exceptional advances. Researches in
this do- main are motivated by the fact that spectral
theory provides a powerful tool to explore statistical
properties of stochastic processes. Indeed, these
techniques proved their efficiency in the survey of
several natural phenomena: in signal or image
processing, in econometrics forecasting, astronomy...
This talk concerns an important class of stochastic
processes having infinite second order moments: it is
about symmetric alpha-stable processes. The
stationary case has been widely studied in the
literature. A general concept of spectral
characterisation of non stationary processes is
discussed in the talk of G. W. Peters. He shows that
under a suitable condition on the covariation
additivity, we can characterize the law of the
alpha-stables processes by a complex bimeasure as it is
the case of second order processes. In this talk we
give details about an efficient estimation of the
bi-spectral density (of the bimeasure) in two
different ways. In the first part of the talk, we use
the moments techniques to give an asymptotically
unbiased estimator of the bi-spectral density from the
continuous time observations of the process. We also
study the statistical properties and consistency of
this estimator. In the second part, we enhance the
results concerning the continues time estimations to
give an asymptotically unbiased estimators from
discrete observations of the process. |
|
(Slides)*
(Video)*
*after publishing |
|
|
Time |
10:30 – 12:00pm |
|
Speaker |
Dr. Gareth W. Peters (UCL) |
|
Title |
Spectral Characterization of the Family
of alpha-Stable Processes that Generalize Gaussian Process
Models |
|
(Slides)* (Video)* *after publishing |
|
|
Time |
13:30 – 15:30pm |
|
Speaker |
Prof. Laurent Clavier (Telecom Lille 1) |
|
Title |
Stable processes and wireless
communications |
|
Abstract |
Telecommunications are undergoing a
significant and important revolution: sensor networks,
5G, cognitive radio, etc. In less than a decade it is
not unreasonable to speculate that autonomous devices
will build their own network: they will adapt to new
environments, change their configuration, adapt to
disappearing or appearing nodes; they will have to
track the best communications opportunities and ensure
reliability with strict energy constraints. However a
complete redefinition of communication rules is needed
to face the main foreseen challenges, interference and
adaptability, with a hard energy constraint.We will
show why new mathematical models are needed.
Alpha-stable distributions are one attractive solution
to represent interference. However it raises
challenges and we will discuss some of them: receiver
design, space/time dependence models and channel
capacity. |
|
Time |
16:00 – 17:00pm |
|
Speaker |
Dr. Malcom Egan (Agent Technology Center) |
|
Title |
Vector Quantization Codebook Design for
Multiuser MIMO |
|
Abstract |
In this talk, we focus on wireless
cellular networks where a base station with multiple
antennas transmits to multiple users
simultaneously; known as multiuser MIMO (MU-MIMO).
This setup plays an important role in current
standards and forms the basis for
advanced interference mitigation techniques, such
as network MIMO. A key aspect of MU-MIMO is that
vector channel state measurements are fed back from
each user in order to optimize transmissions from
the base station to the users. In practice, there are
strong constraints on the capacity of the feedback
links between users and the base station. As
such, an important problem is to design vector
codebooks to reduce the effect of errors due to
quantization (i.e. compression). We show how this
problem can be formulated within frame theory and
propose codebooks based on unitary representations of
groups. |
|
Time |
17:00 – 18:00pm |
|
Speaker |
Prof. Tor Andre Myrvoll (SINTEF) |
|
Title |
Applications of non-parametric copulas to
wireless communications |
|
Abstract |
Copulas provide us with a powerful
framework for modeling multivariate dependency
structures. Parametric copulas like those from the
Gaussian or Archimedean family, are well known and
interpretable, but for greater modeling flexibility,
non-parametric copulas can be considered. In this work
we show how non-parametric copulas can be used to
obtain empirical dependency models of fading
characteristics. |
|
Slides |
|
July 15 (Wednesday)
|
Time |
9:00 – 10:00pm |
|
Speaker |
Prof. Yosihiko Ogata (ERI, U. of Tokyo
and ISM, ROIS) |
|
Title |
Point process modelling: Tutorial with
some topics in statistical seismology |
|
Abstract |
The occurrence times of earthquakes can
be considered to be a point process, and suitable
modelling of the conditional intensity functions of
point process is useful for the investigation of
various statistical features in seismic activity. This
talk introduces models and statistical methods for
analysis of seismicity data, which could be also used
for various data in other research fields. Some
further information is included in my home page:
http://www.ism.ac.jp/~ogata/. However, I avoid
duplication with my@topics at STM2014, http://www.ismvideo.org/STM2014/Yosi.html
in the present talk.
|
|
Time |
10:00 – 11:00pm |
|
Speaker |
Prof. Shiyong Zhou (Peking U) |
|
Title |
Earthquake Hazard Prediction Based on
Seismicity Simulation |
|
Abstract |
Seismicity over 10000 years in Western
Sichuan of China has been simulated based on the
mechanical synthetic seismicity model we developed.
And the maximum magnitude and recurrent time of strong
earthquakes in the interested region can be estimated.
According to the analysis of the simulated synthetic
seismic catalogue, the occurrence of strong
earthquakes with M s ≥ 7.0 in the whole region of
Western Sichuan is rather random, very close to the
Poisson process with seismic rate 0.0454/year, which
means it is reasonable to estimate the regional
earthquake risk with Poisson model in Western Sichuan.
However, the occurrence of strong earthquakes with M s
≥ 7.0 on the individual faults of Western Sichuan is
far from Poisson process and could be predicted with a
time-dependent prediction model. The fault interaction
matrices and earthquake transfer possibility matrices
among the faults in Western Sichuan have been
calculated based on the analysis of the simulated
synthetic catalogues. We have also calculated the
static change in Coulomb failure stress(CFS)on one
fault induced by a strong earthquake on another fault
in Western Sichuan to discuss the physical
implications of the earthquake transfer possibility
matrices inferred from the synthetic catalogue. |
|
Time |
11:00 – 12:00pm |
|
Speaker |
Prof. Sung Nok Chiu (Hong Kong Baptist
University) |
|
Title |
Model-free tests for spatial point
patterns |
|
Abstract |
Data in form of point patterns are often
encountered in spatial statistics, e.g. locations of a
certain species of plants and locations of mines. Point
pattern analysis is often carried out by assuming the
given pattern as a realisation of some spatial point
process model, such as the Poisson process or the
Gibbs process, and then applying parametric methods
for model fitting and testing. However,
certain crucial questions that should be addressed
before modeling are general and are model-free, such
as whether the distribution of the model from which
the data are generated is translationally invariant
(stationary) and rotationally invariant (isotropic) or
whether two given patterns (sampling from e.g. a
normal tissue and a cancerous tissue) have the same
distribution. In
this talk we review the existing tests for these
questions, discuss their weaknesses, and propose new
model-free testing procedures that do not suffer from
such weaknesses.
Simulation results and case studies will be
shown to demonstrate the advantages of our proposed
tests over the tests in the literature. |
|
Time |
13:30 – 14:30pm |
|
Speaker |
Prof. Jiancang Zhuang (ISM) |
|
Title |
|
|
Time |
14:30 – 15:30pm |
|
Speaker |
Prof. Zengping Wen (Institute of
Geophysics, China Earthquake Administration) |
|
Title |
Earthquake damage estimation, Seismic
hazard analysis and Earthquake risk analysis |
|
Abstract |
The estimation of probabilistic future
losses is a matter of increasing interest to those
concerned with public administration in
earthquake-prone regions and those who manage large
numbers of buildings, and they are of particular
concern to the insurance and reinsurance companies
which insure those buildings. A lot of effort has been
devoted to the problem of how to devise reliable
estimates, given the large uncertainties in the
pattern of past earthquake occurrence, both in time
and space, and our limited understanding of the
behavior of the vulnerable elements of the built
environment. This presentation introduces earthquake
damage phenomena, engineering characteristics of
strong ground motion. The two components comprising
the basic structure of a loss estimation study are
presented. One component, the seismic hazard analysis,
involves the identification and quantitative
description of earthquakes to be used as a basis for
evaluating losses. The probabilistic seismic hazard
analysis consists four main steps: 1) determine the
seismic provinces or belts where future earthquakes
may occur according to the law of seismicity; 2)
Determine the parameters of seismicity for each
province and source; 3) select the parameters for
seismic hazard, peak acceleration, response spectrum
and/or intensity, and determine their attenuation laws
suitable for the region; 4) determine the seismic
hazard at the site. The second component, the
vulnerability analysis, entails analysis of the
vulnerability of buildings to earthquakes and the
losses that may result from this damage. A key step is
to establish the relationships among intensity ground
shaking, resulting damage, and associated losses,
expressed by means of a damage probability
distribution. The information assemble from these two
components (seismic hazard analysis and vulnerability
analysis) is combined to produce the probable
distribution of losses for all possible earthquake
events in a given time period and thus to determine
the earthquake risk. The earthquake risk may be
measured in terms of expected economic loss, or in
terms of numbers of lives lost or the extent of
physical damage to property, where appropriate
measures of damage are available. The methodology of
probabilistic earthquake risk analysis to calculate of
the probable levels of losses occurring from all sizes
of earthquake over a period of time, e.g. as expected
losses per year or annualized loss estimation is
presented. Risk analysis can be used to estimate
number of buildings destroyed, lives lost and total
financial costs over a given period of time. Keyword: seismic hazard analysis,
vulnerability analysis, earthquake risk analysis |
|
Time |
16:00 – 17:00pm |
|
Speaker |
Prof. Jen-Tzung Chien (National Chiao
Tung U) |
|
Title |
Bayesian source separation |
|
Abstract |
This talk will survey the state-of-art
machine learning approaches for model-based source
separation with applications for speech separation,
instrumental music separation and singing-voice
separation. The traditional separation approaches
assume the static mixing condition which could not
catch the underlying dynamics in source signals and
sensor networks. The uncertainty of system parameters
may not be precisely characterized so that the
robustness against adverse environments was not
guaranteed. The temporal structures in mixing system
as well as source signals may not be properly
captured. The model complexity or the dictionary size
we assume may not be fitted to the true one in source
signals. With the remarkable advances in machine
learning algorithms, the issues of underdetermined
mixtures, nonstationary mixing condition, ill-posed
condition and model regularization have been resolved
by introducing the solutions of nonnegative matrix
factorization, online learning, Gaussian process,
sparse learning, dictionary learning, Bayesian
inference and model selection. This talk will present
how these algorithms are connected and why they work
for source separation particularly in speech and music
applications. |
|
Time |
17:00 – 18:30pm |
|
Location |
Seminar room 1 (on the 3rd
floor, D305) |
July 16 (Thursday)
|
Time |
9:00 – 10:00pm |
|
Speaker |
Prof. Francois Septier (Telecom Lille 1) |
|
Title |
Langevin and Hamiltonian based Sequential
MCMC for Efficient Bayesian Filtering in
High-dimensional Spaces |
|
Abstract |
Nonlinear non-Gaussian state-space models
arise in numerous applications in statistics and
signal processing. In this context, one of the most
successful and popular approximation techniques is the
Sequential Monte Carlo (SMC) algorithm, also known as
particle filtering. Nevertheless, this method tends to
be inefficient when applied to high dimensional
problems. In this work, we focus on another class of
sequential inference methods, namely the Sequential
Markov Chain Monte Carlo (SMCMC) techniques, which
represent a promising alternative to SMC methods.
After providing a unifying framework for the class of
SMCMC approaches, we propose novel efficient
strategies based on the principle of Langevin
diffusion and Hamiltonian dynamics in order to cope
with the increasing number of high-dimensional
applications. Simulation results show that the
proposed algorithms achieve significantly better
performance compared to existing algorithms. (Joint work
with Dr. Gareth Peters -
http://arxiv.org/abs/1504.05715) |
|
Time |
10:00 – 11:00pm |
|
Speaker |
Prof. Norikazu Ikoma (Kyushu Institute of
Technology) |
|
Title |
Deformable Target Tracking over Infrared
Camera Video in Coarse Resolution with Possible
Frame-Out of the target by Particle Filter |
|
Abstract |
Motion estimation of deformable target
over infrared camera video in coarse resolution with
possible frame-out of the target by an elaborated
particle filter has been investigated. State space
modeling with a state consisting of location factor of
the target and its shape factor has been employed,
where the shape factor is a set of pixels over the
image frame with connectivity property, and the
location factor specifies the position of center of
gravity of the set of pixels over the image plane.
The location factor is in real-valued vector form, while the shape
factor is formed by a set of pixels in discrete
coordinates over the digital image. After having
intensity enhancement of the dim image captured by
infrared camera, particle filter algorithm has been
applied with system model specifying the smooth motion
of the target and its deformation, and likelihood
model to evaluate relatively high intensity values in
the shaped region of the target. Note that for the
system model, we have employed a mixture model of
second order difference equation for smooth motion and
uniform distribution over the image plane to cope with
abrupt change of the location due to lack of frame
rate in the video. For the smooth change of the target
shape, we need to utilize a Markov Chain Monte Carlo
move on the set of connected pixels.Keywords: visual
tracking, deformable object, particle filter, infrared
camera, Markov Chain Mote Carlo. |
|
Time |
11:00 – 12:00pm |
|
Speaker |
Prof. Yoshinori Kawasaki (ISM) |
|
Title |
Scale Mixture of Skewed Kalman Filter and
Its Application |
|
Abstract |
Kalman filter is an essential tool to
yield the log-likelihood of linear Gaussian time
series models via prediction error decomposition. One
of its popular use is to decompose an observed time
series into several unobserved component time series.
As an extension of linear state space model, Naveau et
al. (2005) proposed skewed state space model that
allows one of the decomposed time series to follow
some asymmetric distribution. Assumed class of
distribution is called closed skew normal (CSN) which
includes normal distribution as a special case. The
extended algorithm is called skewed Kalman filter
(SKF). Kim et al. (2014) illustrates scale mixture of
SKF to accommodate a skewed and heavy-tailed
distributed component, but their scheme seems to lack
practical guidelines for actual implementation. In
this presentation, we show fixed interval smoothing
algorithm for SKF which is seemingly missed in the
literature and apply it to daily commodity future time
series obtained from Tokyo Commodity Exchange. Next,
we explore a practically feasible algorithm to
estimate scale mixture of SKF especially when the true
mixture structure and mixture rates are unknown. [This
is the joint work with Mr. Daisuke Kurisu, a master
course student at Graduate School of Economics, the
University of Tokyo.] |
|
Slides
Video |
|
|
Time |
13:30 – 15:00pm |
|
Speaker |
Dr. Pavel V. Shevchenko (CSIRO) |
|
Title |
Modelling Annuity Portfolios and
Longevity Risk with Extended CreditRisk+ |
|
Abstract |
Using an extended
version of the credit risk model CreditRisk+, we
develop a flexible framework to estimate stochastic
life tables and to model annuity portfolios, including
actuarial reserves. Deaths are driven by common
stochastic latent risk factors which may be
interpreted as death causes like neoplasm, circulatory
diseases or idiosyncratic components. Our approach
provides an efficient, numerically stable algorithm
for an exact calculation of the one-period loss
distribution where various sources of risk are
considered. As required by many regulators, we can
then derive risk measures for the one-period loss
distribution such as value-at-risk and expected
shortfall. In particular, our model allows stress
testing and, therefore, offers insight into how
certain health scenarios influence annuity payments of
an insurer. Such scenarios may include improvement in
health treatments or better medication. Using publicly
available data, we provide estimation procedures for
model parameters including classical approaches as
well as MCMC methods. We conclude with a real world
example using Australian death data. |
|
Time |
15:30 – 16:30pm |
|
Speaker |
Prof. Guillaume Bagnarosa (ESC Rennes
School of Business) |
|
Title |
About Risk
Neutral Uncertainty |
|
Abstract |
The pricing of numerous financial
derivatives among which exotic options and variance
swaps rests on the hypothesis of complete markets
which relies in turn on the existence of a continuum
of vanilla option prices and more importantly on the
quality of this information. During this talk we will
highlight several micro-structure sources of noise
which lead undoubtedly to measurement errors and
calibration impediments. While industry and academics
generally assume that the available asset bid-ask
spread mid-price corresponds to the asset fair price,
we demonstrate through high frequency data dynamic
analysis how misleading such an assumption is and how
it relates to another significant source of noise
which is arising from the non-synchronicity of the
information flows. Finally, we propose a state space
model to circumvent this ubiquitous measurement
uncertainty and thus improve the quality of the
information used for derivatives pricing and models
calibration. |
|
(Slides)* (Video)* *after publishing |
|
|
Time |
16:30 – 17:30pm |
|
Speaker |
Mr. Matthew Ames (UCL) |
|
Title |
Violations of Uncovered Interest Rate
Parity and International Exchange Rate Dependences |
|
Abstract |
The uncovered interest rate parity puzzle
questions the economic relation existing between short
term interest rate differentials and exchange rates.
One would indeed expect that the differential of
interest rates between two countries should be offset
by an opposite evolution of the exchange rate between
them, hence ruling out any limited risk profit
opportunities. However, it has been shown empirically
that this relation is not holding and accordingly has
led, over the past two decades, to the reinforcement
of a well-known trading strategy in financial markets,
namely the currency carry trade. This paper
investigates how highly leveraged, mass speculator
behaviour affects the dependence structure of currency
returns. We propose a rigorous statistical modelling
approach using two complementary techniques in order
to demonstrate that speculative carry trade volumes
are informative in both the covariance and tail
dependence of high and low interest rate currency
returns, whereas the price based factors previously
suggested in the literature hold little explanatory
power. We add a new feature to the understanding of
the link between the UIP condition and the carry trade
strategy, specifically attributed to the large joint
exchange rate movements in high and low risk
environments. |
| Slides Video | |
|
Time |
17:30 – 18:00pm |
|
Speaker |
Mr. Lucas Tian Huijie (UCL) |
|
Title |
Loss Reserve |
July 17 (Friday)
|
Time |
9:30 – 10:30pm |
|
Speaker |
Prof. Yoshiki Yamagata, Dr. Daisuke
Murakami (National Institute for Environmental
Studies) |
|
Title |
A spatiotemporal analysis of
participatory sensing data gtweetsh and extreme
climate events toward real-time urban risk management |
|
Abstract |
Real-time urban climate monitoring provides useful information that can be utilized to help monitor and adapt to extreme events, including urban heatwaves. Typical approaches to the monitoring of climate data include the acquisition of weather station monitoring and also remote sensing via satellite sensors. However, climate monitoring stations are very often distributed spatially in a sparse manner, and consequently, this has a significant impact on the ability to reveal exposure risks due to extreme climates at an intra-urban scale (e.g., street level). Additionally, such traditional remote sensing data sources are typically not received and analyzed in real-time which is often required for adaptive urban management of climate extremes, such as sudden heatwaves. Fortunately, recent social media, such as Twitter, furnishes real-time and high-resolution spatial information that might be useful for climate condition estimation. The
objective of this study is utilizing geo-tagged tweets
(participatory sensing data) for urban temperature
analysis. We first detect tweets relating heat
(heat-tweets). Then, we study relationships between
monitored temperatures and heat-tweets via a
statistical model framework based on copula modelling
methods. We demonstrate that there are strong
relationships between gheat-tweetsh and temperatures
recorded at an intra-urban scale, that we reveal in
our analysis of Tokyo city and its suburbs.
Subsequently, we then investigate the application of
gheat-tweetsh informing spatio-temporal Gaussian
process interpolation of temperatures as an
application example of gheat-tweetsh. We utilize a
combination of spatially sparse weather monitoring
sensor data, infrequently available MODIS remote
sensing data and spatially and temporally dense lower
quality geo-tagged twitter data. Here, a spatial best
linear unbiased estimation (S-BLUE) technique is
applied. The result suggests that tweets provide some
useful auxiliary information for urban climate
assessment. Lastly, effectiveness of tweets toward a
real-time urban risk management is discussed based on
the analysis of the results. |
|
Time |
10:30 – 12:00pm |
|
Speaker |
Dr. Ido Nevat (Institute for Infocom
Research, A-Star) |
|
Slides
Video |
|
|
Time |
13:30 – 14:30pm |
|
Speaker |
Prof. Taiji Suzuki (Tokyo Institute of
Technology) |
|
Title |
|
| Abstract |
In this talk,
we present new stochastic
optimization methods that are applicable to a wide
range of structured
regularizations. Structured regularization is a useful
statistical tool to deal
with a complicated data structure such as group
sparsity, graphical sparsity,
and low rank tensor structure. The proposed methods
are based on stochastic
optimization techniques and Alternating Direction
Method of Multipliers (ADMM).
ADMM is a general framework for optimizing a composite
function, and has a wide
range of applications. We propose two types of
stochastic variants of ADMM, which
correspond to (a) online stochastic optimization and
(b) stochastic dual
coordinate descent respectively. Both methods require
only one or few sample
observations at each iteration, and are suitable for
large-scale data analysis.
It is shown that the online type method (a) yields the
minimax optimal
convergence rate in an online setting and the
stochastic dual coordinate
descent type method (b) yields exponential convergence
rate in a batch setting.
|
|
Time |
14:30 – 15:30pm |
|
Speaker |
Prof. Kenji Fukumizu (ISM) |
|
Title |
Kernel Mean Particle Filter with
Intractable Likelihoods |
|
Abstract |
We propose a kernel method for filtering
with a state space model under the condition that the
observation model has no explicit density form but
sampling is possible.
The standard sequential Monte Carlo filters are
not applicable, since they require the functional form
of density functions (up to constant). Using the
framework of kernel mean embedding, the proposed
method expresses a distribution with weighted samples,
for which the Gram matrix computation provides a way
of Bayesian computation.
Resampling with the kernel herding is also
implemented to improve the filtering accuracy. The
experimental results with €alpha-stable multivariate
stochastic volatility models show favorable
performance of the proposed method over existing ABC
filter in the setting of intractable likelihood, and
also in the case of tractable likelihood for high
dimensional variables. This work is a collaboration
with Motonobu Kanagawa and Yoshimasa Uematsu. |
|
Slides
Video |
|
|
Time |
16:00 – 17:00pm |
|
Speaker |
Prof. Konstatin Markov (Aizu U.) |
|
Title |
Probabilistic Dynamic Time Warping based
on Gaussian Processes |
|
Abstract |
A big problem of the DTW based
classification systems is the fact that the class model - the "template" which is
usually one of the training samples, does not account
for the with-in class variability. It is reflected by
the DTW scores and causes bigger overlap between
different classes. Here, we propose a new "template"
model based on Gaussian Processes which accounts for
this variability and, thus, improves the inter-class
separability. |
|
Time |
17:00 – 17:30pm |
|
Speaker |
Prof. Tomoko Matsui (ISM) |
|
Title |
Monte Carlo Dynamic Classifier (MCDC)
Tool |
|
Abstract |
A toolbox developed in Matlab for the
estimation, calibration and filtering of very flexible
and general Gaussian Process specified state space
models is introduced. State-space models are widely
used in many areas of science, engineering and
economics to model time series and dynamical systems.
We develop a toolbox for a complete Bayesian approach
to inference and learning (i.e. state estimation and
system identification) on nonlinear nonparametric
state-space models. |
|
(Slides)* (Video)* *after publishing |
|
|
Time |
17:30 – 17:45 |
|
|
Closing |