GRANT NUMBER: GR/K41090 START DATE: 1/8/94 FINISH
DATE: 31/7/97
DURATION: 36 months AMOUNT: £268,869
PROJECT OBJECTIVES
- To investigate and develop the use of neural network-based advanced
associative memory in knowledge manipulation tasks;
- To develop methods for rapid manipulation of knowledge bases in smart
information systems;
- To develop a prototype system supporting high-speed operations on
inexact symbolic data using correlation matrix memory techniques;
- To demonstrate the application of the results in mission systems.
ACHIEVEMENTS
The AURA project has successfully delivered and demonstrated a prototype
pattern matching engine, using binary neural network based associative memory.
An early prototype integrated (software + hardware) system was operational
in March 1997.
Improved prototype hardware using the Presence design became
operational in early-June 1997, and the current integrated system
demonstrator was operational in October 1997.
More specifically, the project has delivered the following major
results:
- The problems of representing data to CMMs to allow fast matching have
been overcome, using a range of techniques including tensor product and
superimposed binding methods.
- A new and highly efficient method (MBI) for extracting data from the
superimposed binary output of the CMM has been developed.
- New methods have been developed for representing rules in multiple CMMs
to overcome and exploit the partial match problem.
- In multiple CMM systems it is difficult to distinguish instances of
partial match from instances of exact match. Methods have been developed
to overcome this problem.
- It has been shown how the use of superimposed data representations can
deal with commutativity and missing data when input to a neural network
based on CMMs.
- It has been shown that CMMs can be implemented using low-cost
technology with very good speed-ups over existing conventional and
specialised computing systems.
- The utility of the methods has been demonstrated across a range of
problems, including pattern classification, data association, database
and address matching.
- A software library for the construction of systems incorporating AURA
methods has been provided. The library is written in C++ and runs on SGI
platforms and has now reached version 3.0.
- An experimental X Window System GUI version of the system has been
constructed to demonstrate the basic functions provided by AURA in an
easily understood way.
- A small hardware test bed system has been constructed. This is based on
an in-house designed VME board with an FPGA chipset performing the core
CMM computations. The board is installed in an SGI challenge machine, for
evaluating new and potential applications.
MOTIVATION AND APPLICATIONS:
There is a critical need to reduce the workload of pilots in future
generations of aircraft, especially at peak times. A major element of the
workload concerns mission management tasks, and efforts are underway at
British Aerospace (and elsewhere) to develop automated mission management
aids. Such aids would free the pilot to undertake tasks needing particular
human skills.
Typically, these automated aids must manipulate knowledge at very high
speeds due to the high volume of data involved, and the rapid response times
required. In the AURA project we have demonstrated a novel architecture,
providing a fresh approach to some key problems in mission management
systems.
In discussions with BAe during the project we have identified several
areas of mission management where the AURA architecture can offer valuable
improvements in knowledge-manipulation capability. Evaluation work in these
areas continues, and we are now exploring other potential applications at BAe
beyond the original scope of mission management systems. BAe has also agreed
(in principle) to support an application for continuation of the work
commenced in the AURA project.
During the project, significant applications of the AURA concept have been
under development in other projects with The Post Office, and with
Glaxo-Wellcome and Silicon Graphics. The latter project concerns matching of
molecular structures during the process of drug design. This involves
searching databases containing in excess of 100,000 molecules to find those
most similar to a "query" molecule.
In summary, the diversity of this multi-sector industrial support is an
indication of the highly generic nature of the technology developed in the
AURA project. The potential for exploitation through industrial take-up or
product development is now a very real prospect in the near to mid-term.
AURA II PROJECT
The original AURA project ended in July 1997, and the AURA work continues
in a new project called AURA II.