An artificial neural network is simulated that shares formal qualitati
ve similarities with the selective attention and generalization defici
ts seen in people with autism. The model is based on neuropathological
studies which suggest that affected individuals have either too few o
r too many neuronal connections in various regions of the brain. In si
mulations where the model was taught to discriminate children with aut
ism from children with mental retardation, having too few simulated ne
uronal connections led to relatively inferior discrimination of the tw
o groups in a training set and, consequently, relatively inferior gene
ralization of the discrimination to a novel test set. Tao many connect
ions produced excellent discrimination but inferior generalization bec
ause of overemphasis on details unique to the training set. It is conc
luded that within the context of the current model, the neuropathologi
cal observations that have been described in the literature are suffic
ient to explain some of the unique pattern recognition and discriminat
ion learning abilities seen in some people with autism as well as thei
r problems with generalization and concept acquisition. The model gene
rates testable hypotheses that have implications for understanding the
pathogenesis, treatment, and phenomenology of autism.