Computational neuroscience seeks to understand brain function through a combination of mathematical and experimental techniques. The computational neuroscience of mental disorders – a relatively new field of research – may provide an avenue to develop and test the hypothesis of Crespi & Badcock (C&B) that autism and psychosis represent opposite poles of human social cognition. Modifications to existing computational models of healthy brain function, in particular, may help to understand how imbalances in neurotransmitter systems give rise to the behavioral symptoms associated with autism and psychosis.

I will discuss how a computational model of category learning and recognition has begun to inform research on both autism and psychosis (Grossberg Reference Grossberg2000a; Grossberg & Seidman Reference Grossberg and Seidman2006). This model hypothesizes that humans recognize objects by matching top-down expectations (subliminal primes or hypotheses) – based on learned categories – against bottom-up representations of sensory features. According to the model, when there is a sufficient match between top-down categories and bottom-up features, we recognize the observed object as belonging to the selected category. Modifications to the learned associations between features and category can take place at this juncture, thereby refining the category expectation. When there is a sufficiently large mismatch between the categories and features, the brain quickly and unconsciously searches through existing category memories for a better-matching category. Should this search fail, the brain then learns a new category, associating the existing cluster of sensory features with the active category. A model parameter called vigilance determines the amount of similarity required to define a match between categories and feature clusters. If vigilance is high, a high degree of similarity between categories and features is required for recognition to occur, whereas the converse is true if vigilance is low.

It has been proposed that vigilance varies continuously in a population of individuals, and that autistic individuals may have an abnormally high vigilance level (Grossberg & Seidman Reference Grossberg and Seidman2006). This abnormally high vigilance level ensures that autistic individuals are inordinately influenced by bottom-up features. High vigilance leads to the formation of highly specific (concrete) categories and prevents the learning of general (abstract) categories. In terms of recognition performance, high vigilance means that autistic individuals require an above-average degree of similarity between categories and features to recognize an object (Klinger & Dawson Reference Klinger and Dawson2001; cf. Molesworth et al. Reference Molesworth, Bowler and Hampton2005). It also implies that such individuals often fail to recognize objects, such as faces, due to trivial variations in sensory features, such as changes in hair-style. On the flip side, high vigilance ensures that autistic individuals very seldom mistakenly recognize an object as belonging to a category to which it does not belong: High vigilance will lead to the formation of new categories for almost any cluster of features. Such abnormalities in category learning may secondarily affect related brain regions involved in conditioning, motivation, cognition, and motor control (Grossberg & Seidman Reference Grossberg and Seidman2006). Thus, according to this modeling framework, many autistic symptoms may arise from abnormalities in brain systems governing basic learning, attention, expectation, and recognition.

The top-down pathways involved in category learning and recognition usually act as subliminal priming signals. This means that top-down signals do not normally activate neurons encoding sensory features, preventing us from experiencing hallucinations during daily life. It has been proposed (Grossberg Reference Grossberg2000a), however, that humans are able to volitionally (willfully) modulate the influence of top-down signals, allowing these signals to become suprathreshold. Such suprathreshold top-down activation in the absence of actual sensory features is hypothesized to occur during normal imagery, such as inner speech. During psychotic episodes, according to this theory (Grossberg Reference Grossberg2000a), top-down signals become suprathreshold due to the loss of volitional modulatory control, leading to hallucinations – a hypothesis for which there exists some empirical support (Aleman et al. Reference Aleman, Bocker, Hijman, de Haan and Kahn2003). The same modeling framework which helps to explain aspects of autism, is therefore able to account for features of psychosis.

Can the autistic and psychotic features in the above computational framework be considered opposite poles of a spectrum? Probably not, strictly speaking. The high vigilance associated with hyperspecific category formation in autism is in some ways orthogonal to the loss of volitional control of top-down signals in psychosis. It remains theoretically possible, for example, for a person to have both high vigilance and impoverished volitional control. Yet the pertinent issue is whether or not the underlying genetic variation across individuals supports such phenotypes. One recent study examining speech discrimination in schizophrenia (Vercammen et al., in press) suggests that hallucinations may actually be associated with lower-than-normal vigilance levels, supporting the diametrical hypothesis of C&B, at least with respect to vigilance control. Detailed computational modeling of the autistic-psychotic spectrum, should it exist, may help to clarify the relationship between parameters underlying the brain's computational machinery and the genetic factors underlying the development of phenotypic variations in human cognition and behavior.