Impact of the characteristics of data sets on incremental learning

Abstract

Little attention has been paid to identifying the characteristics of a data set that provide favorable conditions for the task of incremental learning. In this work, several metrics were used to characterize data sets and identify the characteristics that may influence the trade-off between stability and plasticity. Three metrics are proposed for the evaluation of stability, plasticity and the trade-off between them in incremental techniques. The experiments were carried out using four incremental neural networks, and the results showed that the shape of the class boundary and spatial distribution of the samples have a great influence on this trade-off.