Detailed learning in narrow fields : towards a neural network model of autism

Autism is a developmental disorder in which attention shifting is known to be restricted. Using an artificial neural network model of learning we show how detailed learning in narrow fields develops when attention shifting between different sources of stimuli is restricted by familiarity preference. Our model is based on modified Self-Organizing Maps (SOM) supported by the attention shift mechanism. The novelty seeking and the attention shifting restricted by familiarity preference learning modes are investigated for stimuli of low and high dimensionality which requires different techniques to visualise feature maps. To make learning more biologically plausible we project the stimuli onto a unity hyper-sphere. The distance between a stimulus and a weight vector can now be simply measured by the post-synaptic activities. The modified "dot-product" learning law that keeps evolving weights on the surface of the hyper-sphere has been employed.Validerad; 2003; 20060922 (ysko