|
Invited Speakers
|
Online Learning, Regret Minimization, and Game Theory
[slides]
[video]
Avrim Blum is Professor of Computer Science in
the Department of Computer Science at Carnegie Mellon University.
The first part of his tutorial will discuss adaptive algorithms for making decisions
in uncertain environments (e.g., what route should I take to work if
I have to decide before I know what traffic will like today?) and
connections to central concepts in game theory (e.g., what can we say
about how traffic will behave overall if everyone is adapting their
behavior in such a way?). He will discuss the notions of external and
internal regret, algorithms for "combining expert advice" and
"sleeping experts" problems, algorithms for implicitly specified
problems, and connections to game-theoretic notions of Nash and
correlated equilibria.
The second part of his tutorial will be about some recent work on
learning with similarity functions that are not necessarily legal
kernels. The high-level question here is: if you have a measure of
similarity between data points, how closely related does it have to
be to your classification problem in order to be useful for learning?
|
|
Latent Variable Models for Document Analysis
[slides]
[video]
Wray Buntine is a researcher in the Statistical Machine Learning group
at NICTA's Canberra Laboratory.
He will consider various problems in document analysis (named entity
recognition, natural language parsing, information retrieval),
and look at various probabilistic graphical models and algorithms
for addressing the problem. This will not be an extensive coverage
of information extraction or natural language processing, but
rather a look at some of the theory, methods and practice of
particular cases, including the use of software environments.
|
|
Inference in Graphical Models
[slides]
[video]
Tiberio Caetano is senior researcher at NICTA's Canberra Laboratory,
where he is a member of the Statistical Machine Learning research program.
His short course will cover the basics of inference in graphical
models. It will start by explaining the theory of probabilistic graphical
models, including concepts of conditional independence and factorisation and
how they arise in both Markov random fields and Bayesian Networks. He will
then present the fundamental methods for performing exact probabilistic
inference in such models, which include algorithms like variable
elimination, belief propagation and Junction Trees. He will also briefly
discuss some of the current methods for performing approximate inference
when exact inference is not feasible. Finally, he will illustrate a range of
real problems whose solutions can be formulated as inference in graphical
models.
|
|
Monte Carlo Simulation for Statistical Inference, Model Selection
and Decision Making
[slides]
[video]
Nando de Freitas is Associate professor in the
Department of Computer Science at the University of British Columbia .
The first part of his course will consist of two presentations. In the first
presentation, he will introduce fundamentals of Monte Carlo simulation
for statistical inference, with emphasis on algorithms such as
importance sampling, particle filtering and smoothing for dynamic
models, Markov chain Monte Carlo, Gibbs and Metropolis-Hastings,
blocking and mixtures of MCMC kernels, Monte Carlo EM, sequential Monte
Carlo for static models, auxiliary variable methods (Swedsen-Wang,
hybrid Monte Carlo and slice sampling), and adaptive MCMC. The
algorithms will be illustrated with several examples: image tracking,
robotics, image annotation, probabilistic graphical models, and music
analysis.
The second presentation will target model selection and decision making
problems. He will describe the reversible-jump MCMC algorithm and
illustrate it with application to simple mixture models and nonlinear
regression with an unknown number of basis functions. He will show how to
apply this algorithm to general Markov decision processes (MDPs). The course
will also cover other Monte Carlo simulation methods for partially
observed Markov decision processes (POMDPs) using policy gradients,
common random number generation, and active exploration with Gaussian
processes. An outline to some applications of these methods to robotics
and the design of computer game architectures will be given. The
presentation will end with the problem of Monte Carlo simulation for Bayesian
nonlinear experimental design, with application to financial modeling,
robot exploration, drug treatments, dynamic sensor networks, optimal
measurement and active vision.
|
|
Marcus Hutter is Associate Professor in the RSISE at the Australian
National University in Canberra and NICTA adjunct.
The first part of his tutorial provides a brief overview of the fundamental methods
and applications of statistical machine learning.
The other speakers will detail or built upon this introduction.
Statistical machine learning is concerned with
the development of algorithms and techniques that learn from
observed data by constructing stochastic models that can be used for
making predictions and decisions.
Topics covered include Bayesian inference and maximum likelihood
modeling; regression, classification, density estimation,
clustering, principal component analysis; parametric,
semi-parametric, and non-parametric models; basis functions, neural
networks, kernel methods, and graphical models; deterministic and
stochastic optimization; overfitting, regularization, and
validation.
Machine learning is usually taught as a bunch of methods that can
solve a bunch of problems (see above).
The second part of the tutorial takes a step back and asks about the
foundations of machine learning, in particular the (philosophical)
problem of inductive inference, (Bayesian) statistics, and
artificial intelligence.
It concentrates on principled, unified, and exact methods.
|
|
Contrast Data Mining: Methods and Applications
[slides]
[video]
Ramamohanarao (Rao) Kotagiri is Professor in
the Department of Computer Science and Software Engineering
at the University of Melbourne.
The ability to distinguish, differentiate and contrast between
different datasets is a key objective in data mining. Such an
ability can assist domain experts to understand their data, and can
help in building classification models. His presentation will
introduce the principal techniques for contrasting different types
of data, covering the main dataset varieties such as relational,
sequence, and graph forms of data, clusters, as well as data cubes.
It will also focus on some important real world application areas
that illustrate how mining contrasts is advantageous.
|
|
Dr. Simon Lucey is Systems Scientist in the
Robotics Institute at
Carnegie Mellon University.
In his tutorial he will cover some of the core
fundamentals in vision and demonstrate how they can be interpreted
in terms of machine learning fundamentals. Unbeknownst to most
researchers in the field of machine learning, the fundamentals of
object registration and tracking such as optical flow, interest
descriptors (e.g., SIFT), segmentation and correlation filters are
inherently related to the learning topics of regression,
regularization, graphical models, generative models and
discriminative models. As a result many aspects of vision can be
interpreted as applied forms of learning. From this discussion on
fundamentals we shall also explore advanced topics in object
registration and tracking such as non-rigid object alignment/
tracking and non-rigid structure from motion and how the
application of machine learning is continuing to improve these
technologies.
|
|
Kernel methods and Support Vector Machines
[slides]
[video]
Alex Smola is leader of NICTA's Statistical Machine Learning Program
in Canberra and Professor in the Computer Sciences Laboratory
at the Australian National University.
His tutorial will introduce the main ideas of statistical learning
theory, support vector machines, and kernel feature spaces.
This includes a derivation of the support vector optimization
problem for classification and regression, the v-trick,
various kernels and an overview over applications of kernel methods.
|
|
Introduction to Reinforcement Learning
[slides]
[video]
Csaba Szepesvari is Associate Professor in the
Department of Computing Science at
University of Alberta.
His tutorial will introduce
Reinforcement Learning, that is, learning what actions to take,
and when to take them, so as to optimize long-term performance. This may
involve sacrificing immediate reward to obtain greater reward in the
long-term or just to obtain more information about the environment. The
first part of the tutorial will cover the basics, such as Markov
decision processes, dynamic programming, temporal-difference learning,
Monte Carlo methods, eligibility traces, the role of function
approximation. In the second part we cover some recent developments,
namely policy gradient and second order methods, such as LSPI and the
modified Bellman residual minimization algorithm.
|
|
Dr. Vishwanathan is senior researcher at NICTA's Canberra Laboratory,
where he is a member of the Statistical Machine Learning research program.
The first laboratory on March 7 will feature some hands on experiments with Elefant (http://elefant.developer.nicta.com.au) mainly concentrating on installing, using, and developing machine learning algorithms within the Elefant framework. We will walk through examples of implementing a simple stochastic gradient descent algorithm as a part of this tutorial.
The second session on March 14 will feature hands on experiments with BNRM (Bundle Methods for Regularized Risk Minimization)
(http://users.rsise.anu.edu.au/~chteo/BMRM.html). The emphasis here will be on developing various loss function modules which can then be plugged into the BMRM solver.
CD's and USB sticks containing installation instructions for Elefant
and BMRM will be handed out during the session. Preferably bring your own laptops, although some spare PC's might be made available. Students will also have a chance to interact with the leading developers of Elefant and BMRM.
|
|
