Data Uncertainty Sensitivity Analysis for Reduced Complexity SVM Classifiers

In this paper we investigate experimentally how different sources of uncertainty affect the classification performance of an SVM based binary classifier. Our aim is to find statistically sound methods for controlling the detrimental effects of such sources when a classifier is to be implemented in hardware platforms where severe limitations force designers to allocate power, computation and memory resources carefully. At a first analysis, SVM revealed robust in terms of noise on data, whereas training data scarcity is a problem to be investigated further on