Skip to Main Content
The application of artificial neural networks in the study of psychopathological syndromes has great potential. Several computational models of acquired and developmental disorders, including autism, have been proposed recently. In this paper, we use the framework of self-organizing maps to study several aspects of autism, by modeling abnormalities in the learning process in biologically plausible manners. We then interpret the resulting feature maps with reference to autistic characteristics. The effects of manipulating the physical structure and size of self-organizing maps were measured and compared with the general characteristics of neural growth abnormalities in autistic children. We find no effect on stimuli coverage, but a negative impact on map unfolding, dependant on the intensity of the abnormality, but not the time of onset. We analyze sensory issues by introducing the concept of attention functions, used to model hypersensitivities and hyposensitivities. The issue of focus on details rather than the whole is analyzed through a model in which distant neighbors are explicitly rejected; we show the model may lead to improved coverage of finely-shaped areas or isolated stimuli, but poorer map unfolding. Finally, we consider effects of noisy communication channels on the development of maps, and show a strong sensitivity of both coverage and unfolding of maps.