School of Computing Science
The Artificial Intelligence and
Neural Nets research group is a collection of researchers interested
in a wide area of AI and Neural Net specialities. Our range of
backgrounds widens our expertise and enables us to discuss a variety
of AI problems. Areas of expertise in the group include: Natural
Language Processing, Distributed Memory, Planning and Self-Organising
Maps. The areas of skills range over Genetic Algorithms, Expert
Systems, Logic, feed-forward Neural Networks, and Computational
Mind and Brain Architecture:
by Geetha Abeysinghe and Christian Huyck
This is an EPSRC funded project which aims to develop a
Natural Language Processing system, which can convert a business
process description in free text to an executable formal
notation. The graphical models of the business process produced by
the execution will then be used for verification and further
elicitation of the business process. The work on the project will
begin in late July/early August.
Cognitive Systems by Roman Belavkin
Cognitive Systems is a general group involving:
multidisciplinary collaboration on building cognitive systems.
- Cross Language Information Retrieval by Viviane Orengo and
Digitized text is stored in many languages. When searching for information,
the user may need a document that is in a language in which he is not
fluent. Multi-lingual thesauri and large multi-lingual corpora can be
used to develop tools that can enable a user to search documents in
languages in which he is not fluent.
- Traveling Salesman and
GAs by Ian Mitchell
The technique, essentially, relies on each node on the TSP graph
sending messages to all possible other nodes in parallel, such a
technique exploits the collision of messages. A genetic algorithm
is introduced to optimise the search space. The use of temporal
representations used to solve the TSP has no illegal
representations encoded in the gene and hence repair algorithms
are unnecessary. This paper investigates the exploitation of
message collision, the post collision process and how certain
sequences of events yield a near optimal solution to the TSP.
Cell Assemblies by Chris Huyck
The CANT (Connections, Associations and Network Technology) model
is designed to function like a natural neural system. The basic
idea is derived from Hebb's idea of the Cell Assembly, a
reverberating circuit of neurons which is the neural equivalent of
a concept. The long-term goal of the model is to discover
how CAs work; discover what CAs can do; duplicate
psychological data with CAs.
- Sequence Recognition using Neural Nets by Ian Mitchell and Siri
Given a non-orthogonal training set, recall a sequence from an
ambiguous stimulus. This usually results in a winner-takes-all
approach, whereby each solution competes until the strongest signal
wins. The winner is proposed as the final solution, however how
correct it is depends on its use. Information retrieval rarely
results in a single output and therefore emphasis is being placed
on techniques capable of retrieving multiple memories. At present
many retrieval methods rank their results, however many sequences
can have equivalent ranking i.e. they are all candidate solutions.
It is this problem domain that this project investigates.
- Natural Language Processing: Chris Huyck,
Geetha Abeysinghe, John Platts, Gill Whitney and Viviane Orengo.
This group works
in many areas of NLP including parsing, word sense disambiguation,
use of language in planning, cross linguistic issues, text extraction,
story production, text summarization, and dialogue agents.
The primary work in this area
is currently being done by Geetha, John, Le Than, Keh Kok, and Viviane.
- Neural Nets: Chris Huyck, Ian Mitchell, Siri Bavan and Usama Hasan.
We work on a wide range of neural networks including traditional
feed forward networks, self-organising maps, growing cell structures
and recurrent networks. We try to apply these systems to the problems
appropriate for them including sequences, concept representation,
categorisation and learning financial data. We are also interested
in the fundamental nature of connectionist systems and biological neural
functioning. The primary work in this area is currently being done by
Ian and Chris.
- Planning: Geetha Abeysinghe, Christian Huyck, and Roger Witts.
Planning includes activities which support better procedures in
diagnosing, monitoring, and goal seeking. Modelling aides the
business procedures by identifying; where subprocesses are
repeated through an organisation and where bottlenecks and
redundancies occur, thus enabling organisations to reach their
objectives more efficiently. We try to apply Natural Language
Processing techniques to process descriptions.
- Describing Mathematics Alexei Vernitski.
Computers have no difficulties in handling mathematics at
mechanical level of logic whereas humans generally work at a much
higher and generally less precise cognitive level. The rigour of
the computer is desirable but without the tedium of lengthy
symbolic logic. We are looking to achieve good human interfaces to
mathematics on computers by providing methods of rigorously
translating between human and computer compatible
representations. There is also the issue of how a human might
navigate to the appropriate level of logical detail or even what
different "levels of logical detail" might mean.
- Genetic Algorithms: Ian Mitchell and Paul Cairns
Investigating the effectiveness of finding optimal solutions to
hard problems, such as maximal clique and the traveling salesman
problem using Genetic Algorithms, GAs. We are currently
investigating how to measure the effectivenes of finding the
solution. Coming up with a metric to measure the effectiveness of a
variant GA often depends on the problem domain, however, if a
representation that is independent of GAs is possible then this
representation can be used as a test bed for future variants of the
- Categorisation: all.
We are hoping to concentrate some effort
on categorisation. Many of us have different approaches to categorisation
including different types of neural networks, symbolic classification
and statistical methods. We are hoping to combine this research to
find fundamental concepts in categorisation. Currently, this work
is in its infancy, and we are trying to find good metrics for measuring
quality of categorisation.
- Text Mining: all.
In the 2001 academic year we are hoping to combine our interests in
categorisation and Natural Language Processing and work on Text Mining.
Middlesex Research Group's Page
- The group meets monthly.
AI reading group has been meeting for almost 5 years.
- The AI group does most of the organisation for Middlesex's