Speaker
Description
Perceptual hallucinations and delusional beliefs are increasingly conceptualized as failures in Bayesian inference, resulting from aberrant weighting of top-down priors relative to bottom-up sensory data. Aberrant precision-weighting is hypothesized to be imbalanced in various psychiatric disorders. To test specific hypotheses about the emergence of hallucinatory and delusion-like behavior in the visual system arising from this mechanism, we implemented a four-layer hierarchical predictive coding (HPC) model, trained on MNIST digits.
Our main findings in this specific architecture include: (1) Dissociation between categorical belief and sensory reconstruction. Confident false top-down prior overrides classification while perception stays faithful, akin to delusional thinking; (2) Convergence and stability of inference are sensitive to changes in precisions of different layers (tested both in isolation and in interactions); (3) Weakening bottom-up trust improves reconstruction accuracy far more under noise (24.7% error reduction) - where the signal is degraded but still present - than under occlusion (4.7%), where information is completely missing; (4) A possible mechanism for relaxing entrenched prior beliefs is reducing a captured false belief’s precision.
This architecture and experimental setup can be adopted as a useful exploratory and explanatory framework for the phenomena arising from the interaction of priors and precisions in HPC models.