Abstract
Viewing situations in which visual input is perceptually ambiguous and consistent with two or more perceptual interpretations typically give rise to bi- or multistable perception, the spontaneous fluctuation between two or more interpretations of the same sensory input. A well-established finding from neuroimaging studies into multistable perception is that spontaneous transitions between perceptual states are associated with neural activity increases in fronto-parietal areas, but it has remained controversial whether such activations reflect cause or consequence of the perceptual transition. Here, we present an account of multistable perception in the computational framework of predictive coding, according to which predictions encoded at higher hierarchical levels are compared against the sensory data represented at lower levels. A mismatch between predictions and sensory data elicits a prediction error signal that is in turn used to update predictions at higher levels. We hypothesized in the framework of predictive coding that fronto-parietal activations may reflect a prediction error signal that is evoked by the currently suppressed percept and that is building up to a point that culminates in a perceptual transition. We used a Bayesian model to estimate the time course of prediction errors during bistable motion perception. Data simulations revealed close similarities between the model's predictions and known temporal characteristics of multistable perception. Fitting the model to behavioural data from an fMRI experiment revealed that prediction error time courses are correlated with neural activity in the inferior frontal cortex, a region that has previously been implicated in causing perceptual transitions. Taken together, we provide theoretical, behavioural and neural evidence for a predictive coding account of multistable perception that posits a crucial role for prediction errors in perceptual inference from ambiguous stimuli.
Meeting abstract presented at VSS 2017