The Neural Correlates of Hierarchical Predictions for Perceptual Decisions

Abstract

Sensory information is inherently noisy, sparse, and ambiguous. In contrast, visual experience is usually clear, detailed, and stable. Bayesian theories of perception resolve this discrepancy by assuming that prior knowledge about the causes underlying sensory stimulation actively shapes perceptual decisions. The CNS is believed to entertain a generative model aligned to dynamic changes in the hierarchical states of our volatile sensory environment. Here, we used model-based fMRI to study the neural correlates of the dynamic updating of hierarchically structured predictions in male and female human observers. We devised a crossmodal associative learning task with covertly interspersed ambiguous trials in which participants engaged in hierarchical learning based on changing contingencies between auditory cues and visual targets. By inverting a Bayesian model of perceptual inference, we estimated individual hierarchical predictions, which significantly biased perceptual decisions under ambiguity. Although "high-level" predictions about the cue-target contingency correlated with activity in supramodal regions such as orbitofrontal cortex and hippocampus, dynamic "low-level" predictions about the conditional target probabilities were associated with activity in retinotopic visual cortex. Our results suggest that our CNS updates distinct representations of hierarchical predictions that continuously affect perceptual decisions in a dynamically changing environment. Bayesian theories posit that our brain entertains a generative model to provide hierarchical predictions regarding the causes of sensory information. Here, we use behavioral modeling and fMRI to study the neural underpinnings of such hierarchical predictions. We show that "high-level" predictions about the strength of dynamic cue-target contingencies during crossmodal associative learning correlate with activity in orbitofrontal cortex and the hippocampus, whereas "low-level" conditional target probabilities were reflected in retinotopic visual cortex. Our findings empirically corroborate theorizations on the role of hierarchical predictions in visual perception and contribute substantially to a longstanding debate on the link between sensory predictions and orbitofrontal or hippocampal activity. Our work fundamentally advances the mechanistic understanding of perceptual inference in the human brain.

Hallucinations and Strong Priors

“…We observe similar hierarchical organization of beliefs relevant to hallucinations in computational analyses of human behavior [34]. In neuroimaging data, high-level predictions about multi-sensory learned associations correlated with activity in supramodal regions such as orbitofrontal cortex and hippocampus [110], while dynamic low-level predictions were associated with activity in primary cortices [110]. Both levels were implicated in hallucinatory perception, but clinical AVHs were associated with dysfunction in higher-level neural and behavioral responses [34].…”

Ketamine Affects Prediction Errors about Statistical Regularities: A Computational Single-Trial Analysis of the Mismatch Negativity

“…To describe a participant's perceptual inference and learning during this roving MMN paradigm, we use a multivariate version of the Hierarchical Gaussian Filter (HGF), a generic Bayesian model introduced by (30) that has been applied in various contexts, such as associative learning (31,45), social learning (32,46), spatial attention (47), or visual discrimination (48).…”

Mind-wandering in Parkinson’s disease hallucinations reflects primary visual and default network coupling

“…These regions include the orbitofrontal and medial prefrontal cortices, which use early low spatial frequency information to generate expectations that constrain ongoing visual processing (Bar, 2003;Bar et al, 2006;Summerfield et al, 2006;Kveraga et al, 2007;Chaumon et al, 2014); hippocampal pattern completion mechanisms that supply memory-based expectations to the visual cortex (Hindy et al, 2016); parahippocampal and retrosplenial cortices supporting the rapid activation of contextual associations during visual processing (Kveraga et al, 2011;Aminoff et al, 2013); and, distinct populations of neurons in the inferior temporal cortex that encode predictions and prediction errors in relation to incoming visual input (Bell et al, 2016;Kok, 2016). The temporal properties of supramodal regions such as the default network are also consistent with the unfolding of high-level predictions over longer timescales (Margulies et al, 2016;Baldassano et al, 2017;Weilnhammer et al, 2018), which fits well with the complex and temporally extended hallucinations that occur in Parkinson's disease (Ffytche et al, 2017). Yet, despite a number of established routes by which the default network may influence visual perception, and evidence of unconstrained default network activity in Parkinson's disease visual hallucinations, we know very little about the behavioural consequences of an over-engaged default network and how this might contribute to hallucinations.…”