Soft-clustering and improved stability for adaptive resonance theory neural networks
Louis Massey · 2008
Stability and plasticity in learning systems are both equally essential, but achieving stability and plasticity simultaneouslyisdifficult.Adaptiveresonancetheory(ART) neural networks are known for their plastic and stable lear- ningofcategories,henceprovidingananswertothesocalled stability-plasticity dilemma. However, it has been demons- trated recently that contrary to general belief, ART stability is not possible with infinite streaming data. In this paper, we present an improved stabilization strategy for ART neu- ral networks that does not suffer from this problem and that produces a soft-clustering solution as a positive side effect. Experimental results in a task of text clustering demonstrate that the new stabilization strategy works well, but with a slight loss in clustering quality compared to the traditional approach. For real-life intelligent applications in which infi- nitestreamingdataisgenerated,thestableandsoft-clustering solutionobtainedwithourapproachmorethanoutweighsthe small loss in quality.