BEGIN:VCALENDAR
VERSION:2.0
PRODID:talks.ox.ac.uk
BEGIN:VEVENT
SUMMARY:New graphical Frameworks for Causal Discovery and Control - Profes
sor Jim Smith (University of Warwick)
DTSTART;VALUE=DATE-TIME:20151015T141500
DTEND;VALUE=DATE-TIME:20151015T151500
UID:https://talks.ox.ac.uk/talks/id/77dafa18-664a-4dfa-98ef-3555c2d8dd71/
DESCRIPTION:Twenty years ago Pearl suggested using the framework of Bayesi
an Networks to formally define collections of causal hypotheses in terms o
f hypotheses about the effects of applying a control. The same semantics h
ad also been proposed by Spirtes Glymour and Sheins to guide the search f
or putative causal hypotheses between high dimensional multivariate obser
vational data. Although this work was seminal in its time I will argue her
e that the proposed semantics certainly do not underpin a general definiti
on of causality and are not entiely applicable to many scientific domains.
In particular both its suppression of the most obvious of axioms - that a
cause happens before an effect and its requirement that a "cause" be asso
ciated with a random variable are each problematic. I will propose various
alternative expressions of causal hypotheses more expressive of the actua
l causal hypotheses than their BN analogues when used in certain contexts:
especially dynamic ones. I will give illustrations from applications asso
ciated with fMRI imaging\, the analysis of gene expression data & longitud
inal behavioral studies.\n\n\n\nSpeakers:\nProfessor Jim Smith (University
of Warwick)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics\,
1 South Parks Road)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/77dafa18-664a-4dfa-98ef-3555c2d8dd71/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:New graphical Frameworks for Causal Discovery and Control
- Professor Jim Smith (University of Warwick)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Distributed Inference (joint work with K. Bleakley and B. Cadre) -
Prof Gerard Biau (Université Pierre et Marie Curie)
DTSTART;VALUE=DATE-TIME:20151022T141500
DTEND;VALUE=DATE-TIME:20151022T151500
UID:https://talks.ox.ac.uk/talks/id/984bc954-5b0d-4832-b7b3-5df9022dfb2d/
DESCRIPTION:The statistical analysis of massive and complex data sets will
require the development of algorithms that depend on distributed computin
g and collaborative inference. Inspired by this\, we propose a collaborati
ve framework that aims to estimate the unknown mean $\\theta$ of a random
variable $X$. In the model we present\, a certain number of calculation un
its\, distributed across a communication network represented by a graph\,
participate in the estimation of $\\theta$ by sequentially receiving indep
endent data from $X$ while exchanging messages via a stochastic matrix $A$
defined over the graph.\n\nWe give precise conditions on the matrix $A$ u
nder which the statistical precision of the individual units is comparable
to that of a (gold standard) virtual centralized estimate\, even though e
ach unit does not have access to all of the data. We show in particular th
e fundamental role played by both the non-trivial eigenvalues of $A$ and t
he Ramanujan class of expander graphs\, which provide remarkable performan
ce for moderate algorithmic cost. \n\n\n\nSpeakers:\nProf Gerard Biau (Uni
versité Pierre et Marie Curie)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/984bc954-5b0d-4832-b7b3-5df9022dfb2d/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Distributed Inference (joint work with K. Bleakley and B.
Cadre) - Prof Gerard Biau (Université Pierre et Marie Curie)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Variance components and residual likelihood - Prof Peter McCullagh
(Department of Statistics\, University of Chicago)
DTSTART;VALUE=DATE-TIME:20151029T141500Z
DTEND;VALUE=DATE-TIME:20151029T151500Z
UID:https://talks.ox.ac.uk/talks/id/021d6506-e689-4677-af2e-a3fe21177dc2/
DESCRIPTION:The marginal likelihood based on residuals in a linear model i
s called the residual likelihood\, sometimes abbreviated to REML. I will d
iscuss the use of REML in a number of applications from animal behaviour\,
growth curves\, agricultural field studies and environmental monitoring.
All of these applications involve designs having several variance compon
ents\, each associated with an independent source of variation such as tem
poral\, spatial or other factors.\n\n\nSpeakers:\nProf Peter McCullagh (De
partment of Statistics\, University of Chicago)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/021d6506-e689-4677-af2e-a3fe21177dc2/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Variance components and residual likelihood - Prof Peter
McCullagh (Department of Statistics\, University of Chicago)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Detecting (multiple) change-points in panel data - Dr Haeron Cho (
School of Mathematics\, University of Bristol)
DTSTART;VALUE=DATE-TIME:20151105T141500Z
DTEND;VALUE=DATE-TIME:20151105T151500Z
UID:https://talks.ox.ac.uk/talks/id/9a88c2b8-182b-4975-8324-b480bbb924a7/
DESCRIPTION:In this talk\, we propose a method for detecting multiple chan
ge-points in the mean of high-dimensional panel data. CUSUM statistics hav
e been widely adopted for change-point detection in both univariate and mu
ltivariate data. For the latter\, it is of particular interest to exploit
the cross-sectional structure and achieve simultaneous change-point detect
ion across the panel\, by searching for change-points from the aggregation
of multiple series of CUSUM statistics\, each of which is computed on a s
ingle series of the panel data.\n\nThe double CUSUM statistic is proposed
as a determined effort for achieving consistency in detecting and locating
(possibly multiple) change-points in the panel data. Its efficiency in ch
ange-point detection is investigated in terms of the cross-sectional size
of the change\, the unbalancedness of change-point location and within-ser
ies and cross-sectional correlations in the panel data. Also\, a comparati
ve simulation study is conducted where the proposed method is applied to a
range of change-point scenarios along with the state-of-the-art.\n\nSpeak
ers:\nDr Haeron Cho (School of Mathematics\, University of Bristol)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/9a88c2b8-182b-4975-8324-b480bbb924a7/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Detecting (multiple) change-points in panel data - Dr Hae
ron Cho (School of Mathematics\, University of Bristol)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Exploiting Symmetries to Construct Efficient MCMC Algorithms With
an Application to SLAM - Prof Csaba Szepesvari (Department of Computing Sc
ience\, University of Alberta)
DTSTART;VALUE=DATE-TIME:20151111T130000Z
DTEND;VALUE=DATE-TIME:20151111T140000Z
UID:https://talks.ox.ac.uk/talks/id/eb866ef8-ff0c-4467-85dd-0809c58ae39c/
DESCRIPTION:Sampling from a given target distribution in an efficient mann
er is a widely studied problem with applications in computer science\, ope
rations research\, statistics\, and many applied subjects. Due to its gene
rality and flexibility\, one of the most successful approach to design eff
icient sampling methods uses the so-called Monte Carlo Markov Chain techni
que\, is the Metropolis-Hastings (MH) algorithm\, which\, in one venue was
named as one of the "top 10" algorithms in computer science. In this talk
we will explore how group moves can be added to MH in general state space
s\, broadening further the applicability of MH beyond what is available to
day. The main motivation for adding group moves to MH is because they allo
w a convenient way to exploit invariances in the target distribution\, eve
n if those only concern a subset of the factors\, or even if they are only
approximate. The main technical difficulty in applying MH in this setting
is the computation of the acceptance probability\, which we address with
tools of topological group theory based on a general result of Luke Tierne
y. The method is demonstrated in an application to robotics in the so-call
ed simultaneous localization and mapping (SLAM) problem where a robot navi
gating a 2D environment and equipped with range only sensors has to learn
a map of its environment while simultaneously estimating its positions. Ou
r experiments with real-world benchmark data on this problem shows that ou
r general method performs competitively with special-purpose algorithms.\n
\nThis is joint work with Roshan Shariff and Andras Gyorgy. The talk is ba
sed on this paper\, that appeared at AISTAT 2015.\n\nSpeakers:\nProf Csaba
Szepesvari (Department of Computing Science\, University of Alberta)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/eb866ef8-ff0c-4467-85dd-0809c58ae39c/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Exploiting Symmetries to Construct Efficient MCMC Algorit
hms With an Application to SLAM - Prof Csaba Szepesvari (Department of Com
puting Science\, University of Alberta)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bayesian nonparametric ordination for the analysis of microbial co
mmunities - Dr Sergio Bacallado (Statistical Laboratory\, Cambridge Univer
sity)
DTSTART;VALUE=DATE-TIME:20151119T141500Z
DTEND;VALUE=DATE-TIME:20151119T151500Z
UID:https://talks.ox.ac.uk/talks/id/54f9fd84-dc45-4111-acef-2eb9c8dbdb4b/
DESCRIPTION:Next generation sequencing has transformed the study of microb
ial ecology. Through the availability of cheap efficient amplification and
sequencing\, taxonomic marker genes such as 16S rRNA are used to provide
inventories of bacteria in many different environments. In particular\, st
udies of the gut microbiome have the potential to shed light on important
health disorders such as obesity\, diabetes\, and Crohn's disease.\n\nWe i
ntroduce a Bayesian factor analysis for discrete samples of species from m
any environments. The marginal prior on the distribution of species in eac
h environment is a normalized completely random measure\, and the dependen
ce between environments is described through latent continuous factors. Th
e procedure is nonparametric in two ways. The number of species is not nec
essarily assumed finite\, and the dimensionality of the factors is learned
from the data. We demonstrate that the analysis yields good estimates of
the distributions of species. We also develop a method to visualize credib
le regions in popular ordination methods applied in microbiology by alignm
ent of posterior samples through conjoint analysis.\n\nJoint work with Boy
u Ren\, Susan Holmes\, Lorenzo Trippa\, and Stefano Favaro.\nSpeakers:\nDr
Sergio Bacallado (Statistical Laboratory\, Cambridge University)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/54f9fd84-dc45-4111-acef-2eb9c8dbdb4b/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Bayesian nonparametric ordination for the analysis of mic
robial communities - Dr Sergio Bacallado (Statistical Laboratory\, Cambrid
ge University)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:t.b.c. - Prof Martin Wainwright (University of California at Berke
ley)
DTSTART;VALUE=DATE-TIME:20151119T154500Z
DTEND;VALUE=DATE-TIME:20151119T164500Z
UID:https://talks.ox.ac.uk/talks/id/3e4f8fd8-21ce-4dda-a0f1-0efb1a7f0493/
DESCRIPTION:t.b.c.\nSpeakers:\nProf Martin Wainwright (University of Calif
ornia at Berkeley)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/3e4f8fd8-21ce-4dda-a0f1-0efb1a7f0493/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:t.b.c. - Prof Martin Wainwright (University of California
at Berkeley)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Regret bounds for Narendra-Shapiro bandit algorithms - Sebastian G
adat (Toulouse School of Economics)
DTSTART;VALUE=DATE-TIME:20150430T141500
DTEND;VALUE=DATE-TIME:20150430T151500
UID:https://talks.ox.ac.uk/talks/id/46db0103-4007-4026-a3a9-637211181706/
DESCRIPTION:Narendra-Shapiro (NS) algorithms are bandit-type algorithms in
troduced in the sixties (with a view to applications in Psychology or lear
ning automata)\, whose convergence has been intensively studied in the sto
chastic algorithm literature. In this talk\, we study the efficiency of th
ese bandit algorithms from a regret point of view. We show that some compe
titive bounds can be obtained for such algorithms in a modified penalized
version. Up to an over-penalization modification\, the pseudo-regret Rn re
lated to the penalized two-armed bandit is uniformly bounded by C sqrt(n)
(for a known C). We also generalize existing convergence and rates of conv
ergence results to the multi-armed case of the over-penalized bandit algor
ithm\, including the convergence toward the invariant measure of a Piecewi
se Deterministic Markov Process (PDMP) after a suitable renormalization. F
inally\, ergodic properties of this PDMP are given in the multi-armed case
.\nSpeakers:\nSebastian Gadat (Toulouse School of Economics)
LOCATION:1 South Parks Road (Lecture Theatre)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/46db0103-4007-4026-a3a9-637211181706/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Regret bounds for Narendra-Shapiro bandit algorithms - Se
bastian Gadat (Toulouse School of Economics)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Scaling Changepoint Detection to Big Data - Paul Fearnhead (Depart
ment of Maths and Statistics\, Lancaster University)
DTSTART;VALUE=DATE-TIME:20150611T141500
DTEND;VALUE=DATE-TIME:20150611T151500
UID:https://talks.ox.ac.uk/talks/id/056999e5-004a-4459-b532-5b27fb1fd604/
DESCRIPTION:Changepoint detection is an increasingly important problem in
a range of applications\, for example to detect copy number variants. A co
mmon approach to inferring the number and position of the changepoints is
to introduce a model for the data within a segment\, and then maximise a p
enalised likelihood function. This maximisation can often be done exactly
using dynamic programming\, but the resulting algorithm has a computationa
l cost that is quadratic\, or even cubic\, in the number of data points.\n
\nThis talk will cover some recent algorithms that can maximise the penali
sed likelihood function exactly\, but at a much lower computational cost.
This includes the first such algorithm that can be shown\, for certain mod
els\, to have an expected computational cost that is linear in the amount
of data.\nSpeakers:\nPaul Fearnhead (Department of Maths and Statistics\,
Lancaster University)
LOCATION:1 South Parks Road (Lecture Theatre)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/056999e5-004a-4459-b532-5b27fb1fd604/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Scaling Changepoint Detection to Big Data - Paul Fearnhea
d (Department of Maths and Statistics\, Lancaster University)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Confidence intervals for high-dimensional Cox model - Yi Yu (Stati
stical Laboratory\, University of Cambridge)
DTSTART;VALUE=DATE-TIME:20150507T141500
DTEND;VALUE=DATE-TIME:20150507T151500
UID:https://talks.ox.ac.uk/talks/id/a54fcb4b-5c91-44e5-88d6-2e7a66dc8256/
DESCRIPTION:The purpose of this work is to construct confidence intervals
for high-dimensional Cox proportional hazards regression models\, where th
e number of time-dependent covariates can be larger than the sample size.
The definition of the one-step estimator is similar to those in van de Gee
r et al. (2014) and Zhang and Zhang (2014)\, but since in the Cox regressi
on model\, the Hessian matrix is based on time-dependent covariates in cen
sored risk sets\, the technical difficulties are fundamentally different.
I will talk about the related theoretical and numerical results in this ta
lk. This is joint work with Jelena Bradic (UCSD) and Richard Samworth (Cam
bridge).\nSpeakers:\nYi Yu (Statistical Laboratory\, University of Cambrid
ge)
LOCATION:1 South Parks Road (Lecture Theatre)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/a54fcb4b-5c91-44e5-88d6-2e7a66dc8256/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Confidence intervals for high-dimensional Cox model - Yi
Yu (Statistical Laboratory\, University of Cambridge)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bayesian network model selection using integer programming - Dr Ja
mes Cussens (University of York)
DTSTART;VALUE=DATE-TIME:20150604T141500
DTEND;VALUE=DATE-TIME:20150604T151500
UID:https://talks.ox.ac.uk/talks/id/36ca2afe-0a29-4e48-8db1-231999afde06/
DESCRIPTION:With complete data and appropriately chosen parameter priors t
he problem of finding a Bayesian network with maximal log marginal likelih
ood (LML) becomes a purely discrete problem: search for a directed acyclic
graph (DAG) with maximal LML. We solve this problem of discrete optimisat
ion using integer linear programming (ILP) with the SCIP (Solving Constrai
nt Integer Programming) framework. In many cases this allows us to solve t
he problem: we find a DAG which we know to have maximal LML. Also using IL
P allows prior knowledge\, such as known conditional independence relation
s\, to be expressed as constraints on DAG structure The key to efficient s
olving is to add certain linear constraints ruling out *cyclic* digraphs d
uring the search. I will report on the successes and limitations of this a
pproach and discuss future directions.\n\nSpeakers:\nDr James Cussens (Uni
versity of York)
LOCATION:1 South Parks Road (Lecture Theatre\, Department of Statistics)\,
1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/36ca2afe-0a29-4e48-8db1-231999afde06/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Bayesian network model selection using integer programmin
g - Dr James Cussens (University of York)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Unbiased computations for MCMC-based inference of Gaussian process
covariance parameters - Maurizio Filippone (School of Computing Science\,
University of Glasgow.)
DTSTART;VALUE=DATE-TIME:20150514T141500
DTEND;VALUE=DATE-TIME:20150514T151500
UID:https://talks.ox.ac.uk/talks/id/0c7b779e-73ae-4e32-9bed-7a04dc71a36b/
DESCRIPTION:Probabilistic kernel machines based on Gaussian Processes (GPs
) are popular in several applied domains due to their flexible modelling c
apabilities and interpretability. In applications where quantification of
uncertainty is of primary interest\, it is necessary to accurately charact
erise the posterior distribution over GP covariance parameters.\n\nEmployi
ng standard inference methods would require repeatedly calculating the mar
ginal likelihood. The formidable computational challenge associated with t
his is that the marginal likelihood is only computable in the case of GP m
odels with Gaussian likelihoods applied to datasets with a limited number
of input vectors (a few thousand). For large datasets\, or for GP models w
ith non-Gaussian likelihoods\, it is not possible to compute the marginal
likelihood exactly\, and this has motivated the research community to deve
lop a variety of approximations techniques. Even though such approximation
s make it possible to recover computational tractability\, it is not possi
ble to determine to which extent they affect the characterisation of the p
osterior distribution over GP covariance parameters.\n\nIn this talk\, I w
ill present the work I carried out over the past few years in the directio
n of developing Markov chain Monte Carlo (MCMC)-based inference methods fo
r GP models that do not require the exact calculation of the marginal like
lihood\, but yield samples from the correct posterior distribution over co
variance parameters. These “noisy” MCMC methods rely only on either un
biased estimates of the marginal likelihood or stochastic gradients (unbia
sed estimates of the gradient of the logarithm of the marginal likelihood)
. I will illustrate ways of obtaining these estimates and demonstrate how
they contribute to the development of practical and scalable MCMC methods
to carry out inference of GP covariance parameters. Finally\, I will demon
strate the effectiveness of these MCMC approaches on several benchmark dat
a and on a multiple-class multiple-kernel classification problem with neur
oimaging data.\nSpeakers:\nMaurizio Filippone (School of Computing Science
\, University of Glasgow.)
LOCATION:1 South Parks Road (Lecture Theatre)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/0c7b779e-73ae-4e32-9bed-7a04dc71a36b/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Unbiased computations for MCMC-based inference of Gaussia
n process covariance parameters - Maurizio Filippone (School of Computing
Science\, University of Glasgow.)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Feature Augmentation via Nonparametrics and Selection (FANS) in Hi
gh Dimensional Classification - Yang Feng (Department of Statistics\, Colu
mbia University)
DTSTART;VALUE=DATE-TIME:20150528T141500
DTEND;VALUE=DATE-TIME:20150528T151500
UID:https://talks.ox.ac.uk/talks/id/06c5bd63-c4de-4469-a5b8-c6bd1a0764d6/
DESCRIPTION:We propose a high dimensional classification method that invol
ves nonparametric feature augmentation. Knowing that marginal density rati
os are the most powerful univariate classifiers\, we use the ratio estimat
es to transform the original feature measurements. Subsequently\, penalize
d logistic regression is invoked\, taking as input the newly transformed o
r augmented features. This procedure trains models equipped with local com
plexity and global simplicity\, thereby avoiding the curse of dimensionali
ty while creating a flexible nonlinear decision boundary. The resulting me
thod is called Feature Augmentation via Nonparametrics and Selection (FANS
). We motivate FANS by generalizing the Naive Bayes model\, writing the lo
g ratio of joint densities as a linear combination of those of marginal de
nsities. It is related to generalized additive models\, but has better int
erpretability and computability. Risk bounds are developed for FANS . In n
umerical analysis\, FANS is compared with competing methods\, so as to pro
vide a guideline on its best application domain. Real data analysis demons
trates that FANS performs very competitively on benchmark email spam and g
ene expression data sets. Moreover\, FANS is implemented by an extremely f
ast algorithm through parallel computing.\nSpeakers:\nYang Feng (Departmen
t of Statistics\, Columbia University)
LOCATION:1 South Parks Road (Lecture Theatre)\, 1 South Parks Road OX1 3TG
URL:https://talks.ox.ac.uk/talks/id/06c5bd63-c4de-4469-a5b8-c6bd1a0764d6/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:Feature Augmentation via Nonparametrics and Selection (FA
NS) in High Dimensional Classification - Yang Feng (Department of Statisti
cs\, Columbia University)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
END:VCALENDAR