Generative Connectionist Models of Cognitive Development : Why They Work

Although the reasons for the success of computer simulations of psychology are often difficult to identify, it is possible to make some progress through systematic experimentation. Reasons for the success of cascadecorrelation models of cognitive development are identified in two case studies. Cascade-correlation is a generative neural network model that constructs its own topology as it recruits hidden units. Capturing correct stage sequences in the integration of velocity, time, and distance cues requires a system that grows in computational power while it learns. Static networks are either too weak or too powerful to capture the full range of stages. Simulating the variation and stages in acquisition of the semantics of English personal pronouns requires sensitivity to differing amounts of addressed and non-addressed speech. Just as with children, networks benefit from the opportunity to hear personal pronouns used in exchanges between o ther speakers. Other simulations suggest that it is important for neural networks to be able to abstract regularities from the environment in order to achieve rulelike behavior and to compute unit activations in a continuous manner to simulate perceptual effects.