Integration of graphical rules with adaptive learning of structured information

Abstract

We briefly review the basic concepts underpinning the ad-aptive processing of data structures as outlined in [3]. Then, turning to practical applications of this framework, we argue that stationarity of the computational model is not always desirable. For this reason we in-troduce very briefly our idea on how a priori knowledge on the domain can be expressed in a graphical form, allowing the formal specification of perhaps very complex (i.e., non-stationary) requirements for the struc-tured domain to be treated by a neural network or Bayesian approach. The advantage of the proposed approach is the systematicity in the spe-cification of both the topology and learning propagation of the adopted computational model (i.e., either neural or probabilistic, or even hybrid by combining both of them).