On-Line Pattern Classification with Multiple Neural Network Systems: An Experimental Study

In the field of pattern recognition, researchers have proposed the application of multiple classifiers to the same data set and the combination of the results using some decision algorithm to improve the performance of individual classifiers [1] [2] [3] [4]. The use of a single system for pattern classification hinges on the assumption that the system is able to capture and process all the input features regardless of what the features might be. In cases where the above assumption fails to hold true e.g. the input features might consist of a mixture of syntactic primitives, linguistic variables, continuous, discrete or nominal attributes, presenting all these features to one classifier for it to make a decision is difficult owing to the diverse types of features. Concatenating all the features into a high-dimensional input vector not only will increase computational burden but will also cause accuracy and tractability problems for some classifiers owing to the so-called "curs-of dimensionality".