Details of Grant 

EPSRC Reference:	GR/J89255/01
Title:	ANALYTICAL APPROACHES TO THE NEURAL NET ARCHITECTURE DESIGN
Principal Investigator:	Kittler, Professor J
Other Investigators:	Petrou, Professor M
Researcher Co-Investigators:
Project Partners:
Department:	Sch of Electronics & Physical Sciences
Organisation:	University of Surrey
Scheme:	Standard Research (Pre-FEC)
Starts:	01 April 1994	Ends:	30 September 1997	Value (£):	172,130
EPSRC Research Topic Classifications:
EPSRC Industrial Sector Classifications:
Related Grants:
Panel History:
Summary on Grant Application Form
The aim of the project is to develop methodology for neural network design. The distinguishing feature of the approach being developed is its strong mathematical basis. In essence, the approach relies on mapping analytical solutions to pattern recognition problems onto appropriate ANN structure and connectivity.One of the expected benefits of the design approach is that the corresponding neural net training schemes will overcome the theoretical requirements imposed by brute force function approximation methods.Progress:During the first year of the project the relationship between contextual pattern recognition techniques represented by probabilistic relaxation methods and conventional neural network architectures has been investigated.In particular we have concentrated on a probabilistic relaxation technique that exploits unary and binary relations of the stimuli. We have shown that the relaxation process maps on a Multilayer Perceptron architecture with specific connectivity, number of layers and activation functions.It has further been shown that it is possible to exploit the relationship between the conditional probability distributions of the relational measurements and the neural network weights. The nature of the relationship suggests that the network weights can be determined from the statistical description of these measurement (input stimuli) distributions. The statistical descriptors are obtained during training by means of standard statistical inference techniques. In consequence, even a small number of training patterns is sufficient to obtain a rough estimate of these distributions which can then be converted into reasonable initial estimates of the neural network weights.
Key Findings
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Potential use in non-academic contexts
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Impacts
Description This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk Summary Date Materialised
Sectors submitted by the Researcher
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Project URL:
Further Information:
Organisation Website:	http://www.surrey.ac.uk