Understanding the cognitive mechanisms underlying autistic behavior: a recurrent neural network study

People with autism spectrum disorder are suggested to exhibit atypical perception and differences in cognitive processing. In behavioral studies, however, such differences are often difficult to verify. Apparently, differences in cognitive processing do not always cause an impairment of behavior. To investigate how such a mismatch between cognitive and behavioral level could be explained, we model and evaluate the process of learning to imitate using recurrent neural networks. We systematically adjust learning parameters of the network which are linked to the precision of learning, a factor that might differ between individuals with autism and typically developed individuals. We evaluate the trained networks in terms of task performance (behavioral level) as well as in terms of the structure of the internal representation that emerges during learning (cognitive level). Our findings demonstrate that comparable behavioral network output can be caused by different internal network representations. A less well structured internal representation does not necessarily result in a decline in performance, but can also be associated with good imitation performance. Additionally, we find evidence that well structured internal representations in our setting emerge with an appropriate integration of top-down predictions and bottom-up information processing, a finding which integrates well with theories from developmental psychology.