Detecting uncertainty regions for characterizing classification problems

A mathematical framework for the analysis of critical zones of the input space in a classification problem is introduced. It is based on the definition of uncertainty region, which is the collection of the input patterns whose classification is not certain. Through this definition a characterization of optimal decision functions can be derived. A general method for detecting the uncertainty region in real-world problems is then proposed, whose implementation can vary according to the connectionist model employed. Its application allows to improve the performance of the resulting neural network.