Boosting the performance of weightless neural networks by using a postprocessing transformation of the output scores

The n-tuple classifier-one of the so-called weightless neural networks-has a number of appealing characteristics such as fast training and response times as well as suitability for hardware implementation. It has been demonstrated that the architecture even in its simple form can be competitive with the more widespread neural network algorithms as well as statistical methods. However, the n-tuple based classification approach has also shown rather poor performance in other cases, especially the architecture has had a tendency to deal inadequately with situations where the training population is highly skewed over the classes in question. A previous theoretical analysis of the classification mechanism has revealed that this short coming of the method essentially can be overcome by invoking a post-processing transformation of the output scores. The paper describes how such a transformation can be derived when the mappings of the output scores are restricted to be linear