Prof. Zachary Warren
School of Medicine, Vanderbilt University
Title: Intelligent Use of Intelligent Technology: Towards autonomous ‘social’ systems of meaning
Abstract: There has been growing interest in utilizing intelligent technology for clinical use in autism spectrum disorder (ASD) – from enhancing detection to social intervention. Researchers have hypothesized that technological tools may be particularly promising as intervention mechanisms, given that many with ASD (1) may better understand physical and visual worlds relative to social worlds, (2) respond well to technologically cued feedback, and (3) show intrinsic interests in technology (e.g. robots, VR, etc.). In addition, researchers have documented numerous systemic and resource barriers towards effective identification and treatment that intelligent technology may be able to overcome (e.g. limited specialists, geographic limits, financial/cost barriers). Unfortunately, to date most clinical applications of technology for ASD have measured behavior in response to simple exposure to robots, toys, or screen-based interactions and enacted limited preprogrammed or confederate dependent interactions. If the ultimate goal of intelligent technology is to engage individuals in effective social intervention over time on a larger scale, meaningful ‘social’ technology may require an interactional framework that dynamically responds to small, meaningful behavioral shifts and links these back to ongoing interactions and learning. In this capacity, the development of dynamic, “closed-loop” interactions involving both technology and aspects of in vivo social interaction may be particularly powerful. “Closed-loop” refers to the ability of a technological system to dynamically interact with an individual with ASD in real-time (i.e. intelligent adaptation), as opposed to an “open-loop” system that behaves in a limited, pre-programmed way. This talk will review ongoing adaptive, “closed-loop” technologically mediated learning systems for ASD intervention. This includes the development and study of socially-relevant intelligent robotic learning environments capable of scaffolding early social orienting and joint attention skills in young children, VR systems for gaze-contingent interaction and driving skills in adolescents, intelligent screening technology for infants and toddlers (e.g. multisensory data capture systems, innovative screening tools), and collaborative virtual learning environments across lifespan activities (i.e. interview systems, social skills interventions, motor/academic interventions). The talk will also discuss and highlight the current challenges related to developing pragmatic, beneficial, and generalizable technological intervention systems for the targeted population (e.g. considerations of heterogeneity within and across individuals over time).
Prof. Yukie Nagai
International Research Center for Neurointelligence, The University of Tokyo
Title: Where do social difficulties come from?: Predictive coding account for autism
Abstract: Predictive coding has been proposed as a unified theory for the human brain. It suggests that the brain perceives the environment and acts on it so as to minimize prediction errors. It has been hypothesized that atypical cognitive characteristics of autism spectrum disorder (ASD) are induced by deficits in the prediction ability. Imbalance between top-down prediction and bottom-up sensation leads to the emergence of atypical internal models, resulting in difficulties in social communication.
My talk presents computational studies to investigate the predictive coding theory. We have been designing neural network models based on the theory and examining how modifications in the model parameters, which correspond to neural impairments in ASD, affect learning of the networks. Our experiments showed that a proper balance between top-down prediction and bottom-up sensation led to the emergence of well-structured internal models. Only such networks achieved higher generalization capabilities as observed in typically developed individuals. In contrast, weaker or stronger influence of prediction produced ASD- or ADHD-like behaviors. Neural networks with weaker predictions acquired rote memorization of given tasks whereas networks with stronger predictions tended to fail in achieving the tasks. Both types of networks exhibited poor adaptation capabilities as observed in ASD.
My talk next presents an HMD display that reproduces atypical visual perception in ASD. We revealed hyper- and hypo-sensitivitiess in ASD’s visual perception and modeled its perceptual process based on predictive coding. The HMD allows typically developed individuals to experience what perceptual impairments people with ASD have and how such impairments affect social disabilities through interaction in the environment. For example, perceptual noises in the vision make it difficult to identify other persons and communicate with them. Taken together, these studies suggest that communication difficulties in ASD are a secondary problem caused by a primary impairment in the prediction ability and that diversities in cognitive and social capabilities of ASD are accounted for by different biases in the predictive brain.