Global, local and embedded architectures for multiclass classification with foreign elements rejection: An overview

In the paper we look closely at the issue of contaminated data sets, where apart from proper elements we may have garbage. In a typical scenario, further classification of such data sets is always negatively influenced by garbage elements. Ideally, we would like to remove them from the data set entirely. Garbage elements are called here foreign elements and the task of removing them from the data set is called rejection of foreign elements. The paper is devoted to comparison and analysis of three different models capable to perform classification with rejection of foreign elements. It shall be emphasized that all studied methods are based only on proper patterns and no knowledge about foreign elements is needed to construct them. Hence, the methods we study are truly general and could be applied in many ways and in many problems. The following classification/rejection architectures are considered: global, local, and embedded. We analyze their performance in two aspects: influence of rejection mechanisms on classification and the quality of rejection. Issues are addressed theoretically and empirically in a study of handwritten digits recognition. Results show that the local architecture and the embedded architecture are advantageous, in comparison to the global architecture.