Peter May - Neural Networks






  - Jump Ratings

  - Juvenile Ratings

  - Flat Ratings

  - AW Ratings

  - Flat Profiler

  - Jumps Profiler


  - TrainerForm (Flat)

  - TrainerForm (Jumps)

  - SireForm (Flat)

  - SireForm (Jumps)

  - My Books


  - About

  - Free Page

  - Links

  - Neural Networks


Neural Networks - Summary         

Artificial neural networks have been defined to be intricate systems of simple units which adapt to their environments, and offer a numerical based approach to processing data, similar to data intensive statistical methods, and in contrast to the more traditional (symbolic) programmed systems.

Artificial neural networks comprise a number of interconnected processing elements (nodes) which are analogous to a biological neuron. Numerical weights are associated with the links (connections) between the nodes, which, essentially, constitute the knowledge of the system. These weights are derived during a period of training, where the system is repeatedly presented with a set of examples until some predetermined convergence criterion is achieved. Hence the system is deemed to learn automatically. They are currently being applied to a wide range of problem domains, including stock market forecasting, mortgage underwriting, and food manufacture.


Although, this representational structure facilitates the learning of complex concepts, it also has its drawbacks.  Generating explanations from connectionist systems is not easy since it requires the unravelling of the weighted connections.  In symbolic systems, on the other hand, it is a straightforward matter to generate explanations from the systemís knowledge-base due to its explicit declarative nature.  This inability to understand the concepts learned by the system makes the acceptance of connectionist systems difficult.  For users to be confident in the performance of a system they need to understand how it arrives at its conclusions.  My research has been aimed at developing solutions to this explanation problem.

Please download any of the following documents if you feel they will be of use.  All are in zipped WORD7.0 format.

Analysis of Neural Network Mapping Functions: Generating Evidential Support, PhD thesis.

 Efficient Rule Extraction From Real-Valued Feedforward Neural Networks