- Info
Dr. Simon O'Keefe
Simon O'Keefe's research interests
My research interests are primarily in the area of neural networks and their application. In particular, I have a long-standing
interest in binary neural networks. These are networks of artificial neurons in which the links between nodes have a
weight of either 0 or 1, unlike "traditional" neural networks that use
continuous weights.
Recent work I have been involved in has been the application of neural
networks to data mining of various forms. Most recently, I have looked at
the use of binary neural networks as a basis for identification of unusual
cases in large customer databases, possibly indicating fraud.
Specific areas in
which I would be interested in supervising graduate students are given
below. However, I will consider any interesting idea.
Use of neural networks in data mining.
The use of binary neural networks for
representing classification structures and the limits on the accuracy which
can be achieved under these conditions: The types of classification
structures that might be implemented are decision tables or decision trees.
Given an implementation of such a classifier by a binary neural system,
what are the limits on the accuracy of the system, and how might the
usefulness of the system be maximised by combination with other (perhaps
slower, more expensive but more accurate) systems?
The use of self-organising networks for data mining
Particularly where the
self-organisation is performed by a binary network. Self-organisation as a
principle is exciting because it allows the data modeller/miner to minimise
the influence of preconceptions on the analysis of the data. However,
interpretation of the model is then more complex because of the lack of
assumptions about the structure in the data.
Neural network models of biological system.
Neural networks are based on a particular view of how real neurons operate. We can look at real systems in two ways. On the one hand, we can use the abstract, artificial model to determine properties that the real system must have in order to function. On the other hand, in order to perform the sort of complex tasks that real systems perform we need to look at the structures and organisation present in real systems, to determine what features add to the computational power of the system.
Chemical substrates for neural networks
An exciting recent development elsewhere has been the use of
chemical substrates for computation. Recent published work has included some
suggestions for the construction of single neurons. I would be interested in
supervising projects to improve on the current models.
Unconventional models for computation, and unconventional systems for implementing conventional computation
Computation is generally thought about in a very narrow way, constrained by the way in which models and implementations of computation have developed. At least as interesting are the alternatives, whatever thay may be.