Generalized relevance learning vector quantization

A life-long learning vector quantization approach for interactive learning of multiple categories

“…The only modification is the incremental adding and testing of new prototype nodes, but compared to cLVQ no feature weighting and selection is performed. In contrast to this, the cGRLVQ additionally applies a feature weighting based on the GRLVQ method proposed by Hammer & Villmann (2002). The GRLVQ weighting is based on the distance d corr c to the nearest correctly labeled prototype w kcorr(c) and d incorr c to the nearest prototype w k incorr (c) with incorrect label:…”

“…The used feature selection method enables the cLVQ to separate co-occurring categories and allows a resource efficient representation of categories, which is beneficial for fast interactive and incremental learning of categories. Recently a variant of an embedded feature selection method for LVQ networks was proposed by Kietzmann et al (2008) based on the GRLVQ method (Hammer & Villmann, 2002) which was called iGRLVQ. This method iteratively removes features with small weighting values λ.…”

“…Euclidean distance), where the chosen metric has a strong impact on the generalization performance. To relax this dependency, metrical adaptation methods can be used that individually weight the different feature dimensions as proposed for the generalized relevance learning vector quantization (GRLVQ) (Hammer & Villmann, 2002) algorithm.…”

“…, K are built up incrementally, where K denotes the current number of allocated vectors w. Each w k is attached to a label vector u k , where u k c ∈ {−1, 0, +1} is the model target output for category c, representing positive, negative, and missing label output, respectively. Each cLVQ node w k can therefore represent several categories c. For the category-specific distance computation d c we use a weighted Euclidean distance with specific weight factors λ cf related to the generalized relevance learning vector quantization (GRLVQ) method proposed by Hammer & Villmann (2002):…”

“…Both methods incrementally add a single prototypical node after 75 wrongly categorized training vectors. For the cGRLVQ the feature weighting learning rate Θ λ was set a magnitude smaller than Θ 0 as suggested by Hammer & Villmann (2002). Additionally for a better stability of the cGRLVQ learning process the incremental insertion of nodes was decoupled from the update of the weighting values.…”