Extracting meaning from cascade networks

Cascade networks have advantages over the more familiar layered feedforward neural network architectures in terms of their ability to solve certain problems, and in their automation of the task of specifying the size and topology of network to use. Cascade networks still share the problem of lack of explanatory mechanism, and remain 'black boxes' sometimes mistrusted by end users. The more complex topologies of cascade networks complicates explanation or rule extraction, hence little previous work has been done. The authors extend their technique based on clusters of characteristic input patterns using the advantages of an improved cascade network.