Neural networks Accelerated pattern-matching boosts recognition
systems.
Could a search engine ever "guess" what your are searching for? For
the past 17 years Jim Austin has been investigating neural
computers. He works as professor of neural computation in the
advanced computer architectures group at York University, and is
chief executive of a technology start-up which builds search
engines based on technology developed from his research at the
university.
Austin's work has involved a project called Aura (Advanced
Uncertain Reasoning Architecture), a set of high-speed search
techniques for matching patterns in unstructured data.
"We want to solve the problem of how to get machines to recognise a
large number of objects," said Austin. This involves recalling and
matching sounds, smells, tastes and images quickly. The Aura engine
can be used in areas such as postal address matching, high-speed
rule-matching systems, matching 3D molecular structures and
trademark-database searching.
Austin demonstrated how Aura could be used when he joined Business
Trade International's UK contingency that exhibited at this year's
CeBit show.
Aura is at the heart of Fedaura, a £1.4m project for the Department
for Work and Pensions looking at ways to reduce benefit fraud using
its text-searching engine. It is also being used in an e-science
pilot project called Dame (Distributed Aircraft Maintenance
Environment), where it listens to vibration signatures in aircraft
engines to detect when maintenance is required.
Another project involves graph matching, developed by
GlaxoSmithKline, which could be used for pattern matching in drug
discovery projects.
In normal databases, data can be indexed to allow a user to search
for a given piece of information quickly without having to compare
every item in the database. Clean data is required to construct
such an index, as well as knowledge of how the data is to be
accessed.
For example, unless the index takes into account every spelling
variation of each name, a lookup on the database will fail if the
text in the query is misspelt. If the user has an address and wants
to find the corresponding name, the system would need an additional
index for the address field (which again requires clean data).
Without this, it would be inflexible and have limited
applications.
Computers are good at handling tasks that can be modelled as a
series of processes or calculations, but they are less efficient at
dealing with other types of information processing.
In particular, it is extremely difficult to program computers to
recognise patterns like speech, images or mis-typed text. Neural
computing is a research area that investigates how computers can
solve problems by mimicking the way the brain works through
building neural networks.
"Conventional neural networks are based on mathematics, where the
network stores information as continuous numbers," Austin said. In
his approach, data is held as binary ones and zeros. To achieve the
speed to run the searches, Austin said the code needs to use
low-level computer operations.
Most neural networks rely on learning. Users have to "train" speech
recognition or writing recognition software to understand their
voices or read their handwriting before the software can be
used.
Training is good for personal use, but does not help when speed is
important, as with face recognition or reading a postal address
label, where the machine has just one shot at recognising the
image.
Aura allows a user to compare an unknown item against every item
stored, and retrieve the examples that are the most similar very
quickly, said Austin.
The benefit of this approach is that Austin has been able to build
hardware to speed up searches. Based on field programmable gate
arrays and digital signal processors, the design, called Presence
1, is configured with 4Gbytes of memory and plugs into a Sun or PC
server. The current installation, CortexOne, uses 24 Presence 1
cards to accelerate searches.
Austin has built a network of 24 PC servers configured as a Linux
Beowulf cluster. Each PC in the cluster is equipped with a Presence
1 card, giving a total of 24 cards in operation.
CortexOne was able to run 11 matches per second when used for
postal address matching he said. This month a new version of the
card, Presence 2, will be installed and Austin is expecting a
tenfold increase in speed.
CV: Jim Austin
Jim Austin is professor of neural computation at York
University. He has directed the advanced computer architectures
group for more than 15 years. The group is one of the largest units
in the computer sciences department at York University, with more
than 35 researchers, technicians and academics.
Austin has more than 170 published works, and he has
participated in many research projects involving the likes of
GlaxoSmithKline, Royal Mail, BAE Systems, Rolls Royce and
EDS.
He is most well known for his work on binary neural networks and
the Aura technology. Austin's research is primarily motivated by
neurobiology. His primary interest is neural networks but he also
works on applications for grid computing, computer vision and
advanced computer architectures.
He is chief executive of the recently-established,
privately-funded university spin off Cybula, which is focused on
transferring the Aura technology from the university into industry
applications.
The company operates in conjunction with Austin's group from the
Science Park at the University of York.