A bridge between physiological and perceptual views of autism by means of sampling-based Bayesian inference

ECHEVESTE, RODRIGO; FERRANTE, ENZO; MILONE, D.H.; SAMENGO, INÉS

Berlin

Conferencia; Bernstein Conference 2021; 2021

Theories for autism spectrum disorder (ASD) have been formulated at different levels: ranging from physiological observations to perceptual and behavioural descriptions. Understanding the physiological underpinnings of perceptual traits in ASD remains a significant challenge in the field.Here we studied the link between Bayesian computations under weakened priors, and inhibitory dysfunctions, neural variability, and oscillations in ASD subjects. We worked with a recurrent neural-network (RNN) model recently described in [1], which had been trained to perform inference in a visual task via sampling. This model is able to represent full posterior distributions, and not just point estimates of the stimuli, while displaying characteristic dynamical features of cortical responses, such as gamma oscillations and transient overshoots. Here we took two parallel paths: In one of them, we modified the probabilistic generative model under which the stimuli are assumed to be generated in order to increase the uncertainty of the prior distribution [2]. In the other, we weakened the inhibitory connections in the neural network to induce an inhibitory dysfunction [3]. We found that both paths lead to consistent results in terms of the represented posterior distributions (Fig. 1, cf. panels a-b), providing support for the view that both descriptions might constitute two sides of the same coin.The dynamical properties of the network performing Bayesian inference allowed us to connect the higher uncertainty in the estimation of stimuli that results from weakened priors and the increased neural variability observed in ASD subjects [4,5]. Moreover, the model captures previous experimental observations in the autistic population in terms of gamma oscillation power [6] and peak frequency 7. Finally, we studied peak transient responses in the model and made connections to predictive-coding views of ASD 8.In short, we show how RNNs optimized for sampling-based inference are viable candidates to bridge the gap between Bayesian perceptual theories of ASD and their physiological underpinnings in terms of inhibitory dysfunction, neural variability and oscillations. We believe these results highlight the potential for the use of the emerging body of function-optimized neural networks as models to establish mechanistic links between neural activity and computations in the cortex that go beyond the study of neurotypical perception.