ABSTRACT
Interactive Activation and Competition (IAC) networks have been successfully used as cognitive models yielding results functionally isomorphic with human responses across a wide range of psychological tasks. To date, all of these models have been designed by hand, and the use of learning techniques to modify each model has been severely restricted. This paper demonstrates that the identification of patterns in sub-populations of the input, paired with a statistical analysis of the relationship between each pattern and the remaining inputs can be used to autonomously generate architectures functionally and structurally similar to those used in IAC models. A new hybrid-learning method is introduced. And a comparison between hand-designed and learnt models is provided demonstrating that the new methods preserve the functionality of the original model. The new method is seen as beneficial to the modeler as it completely removes the need to hand design IAC architectures and reduces the number of assumptions made. It is further suggested that the continuous adaptation of these models allows for the analysis of developmental trajectories, thus broadening their applicability to include developmental theories in cognitive science.
