Encoding auditory-visual interactions in a neural model of sensory abnormalities in autism

摘要

Sensory abnormalities and multisensory integration effects play an important role in autism. This paper deals with the latter, using a neural model that suggests that weak central coherence, a processing bias for features and local information, may account for a failure to control or modulate the experience of sensory inputs. A series of algorithms is proposed to encode auditory-visual interactions through `sensitivity functions', which model lateral neural interactions that reinforce or weaken learning in a self-organizing map. The framework developed targets a specific profile, whereby enhanced responses are seen in an individual with autism for high-randomness content in video signals, a tone-like audio spectrum, and a sharp (positive) transition in the entropy of the video signal acting as a trigger. Certain processing blocks can be easily modified to target different profiles, and performance of the algorithms can be tuned by adjusting several parameters. A comprehensive set of audio-visual recordings is used to evaluate performance; in terms of identification of scenes eliciting enhanced responses, the method yields a good false positive ratio (under 10%), and a modest false negative ratio (just under 40%). Possible avenues for improvements and further research are suggested.