Towards a Neurocomputational Model of Sensory Differences in ASD

Background:

ASD is associated with differences in sensory processing, including difficulties organizing sensory experiences (Dakin & Frith, 2005). Predictive processing theories of cognition explain these differences in terms of differences in statistical processing of sensory information in the brain. A core hypothesis of the predictive processing account is that people with ASD rely less on prior perceptual experience, and more on detailed sensory information, when making perceptual inferences about things in their environments (Pellicano & Burr, 2014). This account has been supported in recent experiments (Skewes et al, 2014; Skewes & Gebauer, 2016).

Objectives:

The purpose of the present research is to use computational modeling to pinpoint the specific mechanisms which underlie the tendency in ASD to rely less on prior sensory information.

Methods:

In the experiments cited above, participants are asked to make binary perceptual inferences given continuously distributed stimuli. For instance, in Skewes and Gebauer (2016), participants were presented with sounds which varied in the location of their spatial sources, and were asked to report whether each sound was produced by a “white” cricket, whose territory was distributed more towards the left auditory hemifield, or by a “black” cricket, whose territory was distributed more towards the right. Participants were given feedback at the end of each trial, and were rewarded for correct responses. Participants with ASD integrated prior information about the base-rates of each kind of cricket less optimally.

Drawing resources from Signal Detection Theory and Prospect Theory, we have developed a cognitive model of the processes underlying this integration. We model learning during the task as an interaction between 1) sensitivity to sensory error of judgements, and 2) sensitivity to feedback. This allows us to develop a deeper explanation of the functional differences identified for individuals with ASD.

To understand the neural mechanisms implementing these differences, we’ve adapted a neural network model designed to learn the same kind of perceptual inference required by the task (Helie, 2014). We used the experimental results to constrain simulations of the network. The results of these simulations allow us to make precise predictions about the neurobiological mechanisms underlying the functional differences observed in ASD.

Results:

In a novel pilot experiment designed to test our cognitive model, we found that autistic traits in the neurotypical population predicted sensitivity to feedback, but not sensitivity to sensory error. Based on this result, we adjusted the rate of learning in frontal GABAergic/dopaminergic synapses in the neural network, which represent the rate at which feedback drives learning in the model. Using simulations, we found that when the learning rate was set lower, the network performed in a way similar to participants in the cited studies.

Conclusions:

This research theoretically motivates the hypothesis that sensory differences in ASD result from differences in neurobiological mechanisms in frontal networks involved in learning about how to reduce continuous sensory information to informative perceptual categories.