Ranking Pattern Recognition Features for Neural Networks
In all pattern recognition technologies, a crucial question is what features to extract and use for the basis of a recognition system. There are well-established techniques for ranking features to be used in linear combination, but when non-linear combination technologies, such as neural networks, are contemplated feature significance ranking is more problematic. In this paper we present a practical technique for ranking features in terms of significance for a neural-net pattern recognizer. We provide the results of applying this clamping technique to a small selection of problems, both well-defined abstract problems that permit a precise exploration of the technology, and more realistic data-defined problems that demonstrate its practical worth.
KeywordsNeural Network Input Feature Generalisation Rate Redundant Feature Trained Neural Network
Unable to display preview. Download preview PDF.
- Tabachnick, B. and Fidell, L. “Using Multivariate Statistics” 3rd Edition, HarperCollins College Publishers, New York, 1996.Google Scholar
- Bishop, C. “Neural networks for Pattern Recognition” Claredon Press, Oxford, 1995.Google Scholar
- Le Cun, Y. et al. “Optimal Brain Damage”. in D.S. Touretzky,. ed.Advances in Neural Information Processing Systems,1990, Vol. 2, pp. 598–605.Google Scholar
- Wang, W., Partridge, D. and Rae, S. “Multiversion systems of neural networks for predicting the risk of osteoporosis”, Proc. 11th Internat. Florida AI Research Conference.,Sanibel Island, 18–20 May, 1998, pp. 317–321.Google Scholar
- Partridge, D., Tallis, D. and Wang, W. “Inductive programming and a multiversion approach to data mining”, Proc. IASTED Conf. on AI and Soft Computing,Cancun, Mexico, 27–30th May, 1998, pp. 19–22.Google Scholar