On the characteristics of growing cell structures (GCS) neural network

Abstract

In this paper, a self-developing neural network model, namely the Growing Cell Structures (GCS) is characterized. In GCS each node (or cell) is associated with a local resource counter tau (t). We show that GCS has the conservation property by which the summation of all resource counters always equals s(1 - alpha)/alpha, thereby s is the increment added to tau (t) of the wining node after each input presentation and alpha (0 < alpha < 1.0) is the forgetting (i.e., decay) factor applied to tau (t) of non-wining nodes. The conservation property provides an insight into how GCS can maximize information entropy. The property is also employed to unveil the chain-reaction effect and race-condition which can greatly influence the performance of GCS. We show that GCS can perform better in terms of equi-probable criterion if the resource counters are decayed by a smaller alpha.