|
David Rees
School of Mathematics and Statistics, University of Sydney
Ghost in the machine: From turbulence, faces and ducks to solar magnetic fields
Wednesday 24th, March 14:05-15:55pm,
Carslaw Building Room 359.
In a recent paper in Science (Vol 293, p 2051, 2001), Mjolsness and
DeCoste at the US Jet Propulsion Laboratory emphasized the importance
of machine learning (e.g. principal component analysis - PCA, neural
networks - NNs, and support vector machines - SVMs), for the future of
science. A key driver for such a marriage between science and
information technology is the overwhelming flood of digital data
facing the scientist of the 21st century. Indeed one can argue that
information technology cannot develop in a vacuum, but must be wedded
to specific applications.
Over the last decade, while at CSIRO, I was involved in a number of
projects applying machine learning to signal and image processing,
often using PCA. Thus much of this talk is a ghost story, the ghosts
being the PCA eigenvectors. The story starts in the early 1990s in
CSIRO's Synergetics project which focused on the nexus between pattern
formation and pattern recognition in complex dynamical systems. In
particular, Sirovich's application of PCA to the control of turbulence
and to low dimensional representation of human faces inspired the
development of SQIS (System for Quick Image Search), a system for
detection and recognition of faces in real time video. SQIS is now a
mature technology that is being commercialised. In the mid 1990s
Nayar's group at Columbia University showed how PCA could be used in
robotics, e.g. to estimate the pose of an object for pick and place by
a robotic arm. My favourite object is Nayar?s toy duck because it was
the duck that triggered the idea of using PCA for fast solution of
inverse problems such as the estimation of vector magnetic fields in
the solar atmosphere from measurements of polarization spectra. The
Paris Observatory and the US High Altitude Observatory now use PCA
routinely because (1) it is at least two orders of magnitude faster
than traditional data fitting by least squares, (2) it gives extra
physical insight, and (3) it enables the solution of otherwise
intractable problems such as weak magnetic field measurements in solar
prominences where the quantum theory of line formation is a
frightening mess.
I will conclude with a summary of the latest developments aimed at
real-time inversion of polarimetric data. Recently HAO has implemented
a new inversion method based on NNs. This is much faster than PCA and
looks like it will be the method of choice for on-board data
processing on a satellite to be launched in 2007. But I?m still
holding out hopes for another method developed and patented by
CSIRO. This is called Multiple Support Vector Regression, a generic
inversion method, which still needs considerable research - an ideal
topic for a PhD student in the School of Mathematics and Statistics.
|
|
|
|
|
|
|
|