Designing efficient distributed neural classifiers: application to handwritten digit recognition

Describes an automatic method for building distributed neural classifiers for pattern recognition. The methodology is based on the detection of reliable regions in the representation space, i.e. clusters exclusively composed of patterns from the same class. This detection is performed using a hierarchical clustering method associated with the supervised information provided by a professor. The proposed methodology consists of associating each of these regions with a multilayer perceptron (MLP) which has to recognise elements that are inside its region while rejecting all others. Experimental results for a real problem (handwritten digit recognition) reveal an interesting generalisation behaviour of the distributed classifier in comparison to the k-nearest neighbour algorithm as well as a single MLP