Online Labelling Strategies for Growing Neural Gas

Growing neural gas (GNG) has been successfully applied to unsupervised learning problems. However, GNG-inspired approaches can also be applied to classification problems, provided they are extended with an appropriate labelling function. Most approaches along these lines have so far relied on strategies which label neurons a posteriori, after the training has been completed. As a consequence, such approaches require the training data to be stored until the labelling phase, which runs directly counter to the online nature of GNG. Thus, in order to restore the online property of classification approaches based on GNG, we present an approach in which the labelling is performed online. This online labelling strategy better matches the online nature of GNG where only neurons – but no explicit training examples – are stored. As the main contribution, we show that online labelling strategies do not deteriorate the performance compared to offline labelling strategies.