Cascade Processing

According to functional analysis (Cummins, 1983), a complex task can be broken down into distinct stages of information processing. In many models, these stages can be sequentially ordered, the complex task can be performed by completing each distinct stage in sequence (e.g. Marr, 1982). In cascade modeling, a later stage of information processing can begin operating before the completion of earlier stages.

A variation of cascade processing, called cascade correlation (Fahlman & Lebiere, 1990), has been used to develop connectionist networks. In cascade correlation, a simple network is trained on a task. When the network can no longer improve, a single unit is added to it, and additional training occurs. This process is repeatedly iteratively until enough units have been added to permit the network to learn the desired task. Some researchers in developmental psychology have used the addition of a unit in cascade correlation to simulate a change in stage of cognitive development (e.g. Shultz & Schmidt, 1992).

References:

1. Cummins, R. (1983). The Nature Of Psychological Explanation. Cambridge, MA.: MIT Press.
2. Fahlman, S. E., & Lebiere, C. (1990). The Cascade-Correlation Learning Algorithm (No. CMU-CS-90-100). Pittsburgh, PA: School of Computer Science, Carnegie Mellon University.
3. Marr, D. (1982). Vision. San Francisco, Ca.: W.H. Freeman.
4. Shultz, T. R., & Schmidt, W. C. (1992). A cascade-correlation model of balance scale phenomena. Paper presented at the Thirteenth Annual Conference Of The Cognitive Science Society.

(Revised March 2010)