Automatic growing of a Hopfield style network during training for classification

• Published in: Neural networks 1997-04, Vol.10 (3), p.529-537
• Main Author: BROUWER, R. K
• Format: Article
• Language: English
• Subjects: Applied sciences; Artificial intelligence; Computer science; control theory; systems; Connectionism. Neural networks; Exact sciences and technology
• ISSN: 0893-6080; 1879-2782
• DOI: 10.1016/S0893-6080(96)00087-1

Hopfield networks, a type of Recurrent Neural Network, may be used as a tool for classification by storing exemplars as memories. This method of using the Hopfield network for classification has certain shortcomings, such as the limits on the number of class exemplars that can be stored and the size of the connection matrix required when used for classification of images. This paper describes a method of growing a Hopfield style network for use in classification of patterns. The complete network that is grown consists of three networks in sequence with the middle network being a fully recurrent Hopfield style network. The first network is a one layer feedforward network while the last network is simply a selector network which selects components from the terminal state of the recurrent network. The Hopfield style network grows automatically during training as additional nodes are required. The resulting network can be trained for the purpose of classifying bi-polar vectors. Connection matrices are determined, using a modified Widrow-Hoff learning rule, such that the exemplars are attracted to exemplars or prototypes within the same class. An unclassified element is then classified by the class of its attractor. No pre-processing is required to determine prototypes and all the training elements are used directly. A reduction in the number of arithmetic operations from order of magnitude n super(2) to n takes place by growing the network rather than initialising the network to size n. The method is successfully applied to the classification of cervical cells for cancer detection and to the classification of diabetes patients recorded in the 'Pima Indians Diabetes Data Base'. Results depict how the network grows as learning takes place.