 ast
month, Dr. Lawrence Saul’s work entered the top
1% in terms of total citations in the field of Engineering
in Essential
Science Indicators .
His current record in this field includes 8 papers cited a
total of 304 times. Dr. Saul is an Associate Professor in the
Department of Computer Science and Engineering at the
University of California, San Diego. In the interview below,
he talks with in-cites about his highly cited work.
|
Would you give us some
background on your education and early research?
|
|
|
UCSD Jacobs
School of Engineering |
|
I was not actually trained as a computer scientist. I majored in
physics as an undergraduate, and then I obtained a PhD in Physics.
Like many physicists at the time, though, I developed an interest in
neural networks which eventually led me into the field of machine
learning.
What do you consider the main focus of your research?
I am mostly interested in machine learning and pattern
recognition. For example, how can machines learn from experience to
recognize faces and voices as well as the average toddler does?
Your most-cited paper is the 2000 Science paper,
"Nonlinear dimensionality reduction by locally linear
embedding." Please talk a little about this paper—what is LLE,
what is its significance, and how has it changed the field?
LLE stands for locally linear embedding. It is a method for
analyzing and visualizing high-dimensional data. There are many
linear approaches to this problem—to determine, for example, if a
set of high-dimensional points are concentrated along a
(one-dimensional) line, near a (two-dimensional) plane, or more
generally in a low-dimensional subspace. LLE is a nonlinear approach
to this problem: it can be used even more generally to analyze
high-dimensional points that lie on or near a low dimensional
manifold.
Its significance lies in the fact that the algorithm is nearly as
simple as strictly linear approaches, and yet in many cases it can
reveal much more. At the time that LLE was published, this was
fairly surprising: most researchers did not believe that
optimizations for nonlinear dimensionality reduction could be
anywhere near as tractable as optimizations for linear
dimensionality reduction. Since the publication of the 2000 Science
paper, there has been an explosion of work on this subject. The
field continues to be active, with many new approaches borrowing
from or appealing to the basic ideas of LLE.
Is LLE being applied in face-recognition technology for
security? Is it competing with other such technologies?
|
Larger view |
|
Example Picture of Locally Linear Embedding
|
|
|
An image is simply a collection of pixels; as such, it can be
viewed as a point in a high-dimensional space, with each pixel
representing one of the dimensions. Many researchers in computer
vision have used LLE to preprocess images for various forms of
pattern recognition. I am sure that this has been tried for face
recognition. (I'm not sure, though, that it represents the state of
the art.)
What other practical applications have arisen or are expected
to arise as a result of this research?
Another practical application has been data analysis in
experimental neuroscience. For example, a visualization based on LLE
was shown on the cover of the September 2003 issue of the journal Neuron.
Recently, in a related line of research, we have studied the
problem of localization in large sensor networks. Each node in these
networks can measure, to some degree of accuracy, the distance to
its nearest neighbors. From noisy estimates of these local
distances, the problem is to recover the global geometry of the
network.
Where do you see this field in 5, 10 years?
Currently, the algorithms are used mainly in batch mode, to
analyze previously collected data sets of moderate size. In 5-10
years, we should have faster algorithms capable of analyzing much
larger data sets. We should also be able to work in an online
setting, where the data is being analyzed as it is being collected,
and where the analysis provides real-time feedback that directs the
data acquisition. Hopefully, we will also be able to analyze richer
and more complicated types of data in all areas of science and
engineering.
Dr. Lawrence K. Saul
Department of Computer Science and Engineering
University of California, San Diego
San Diego, CA, USA
| Dr. Lawrence Saul's
most-cited paper with 210 cites to date: |
|
Roweis ST, Saul LK, "Nonlinear dimensionality reduction by locally linear embedding,"
Science 290(5500): 2323-+, 22 December 2000. |
|
Source:
Essential Science Indicators |
|
|
cites
