Modeling Delusions through Hybrid Predictive Coding

Abstract

Delusions, fixed and false beliefs resistant to counterevidence, are hallmark symptoms of psychotic disorders, yet their underlying mechanisms remain poorly understood. This thesis investigates the hypothesis that hybrid predictive coding (HPC) provides a computationally and cognitively plausible framework for modeling delusional belief formation [1]. While standard predictive coding models posit that the brain performs perception and inference through continuous, gradient-based error correction between predictions and sensory inputs [2], the hybrid approach augments this framework by integrating a fast, amortised inference mechanism [3]. This component allows for rapid, feedforward guesses about the causes of sensory input, shaped by previously learned input-cause mappings.

In this work, we hypothesize that delusions may arise when: (1) these amortised mappings are biased or trauma-informed; (2) prior beliefs are assigned high precision, diminishing their regulatory influence; and (3) the iterative component of inference fails to adequately correct faulty initial guesses. This thesis aims to experimentally implement and evaluate such a model, systematically manipulating model parameters to simulate conditions thought to give rise to delusional cognition.

By constructing both “healthy” and “delusion-prone” versions of the model, we simulate scenarios involving ambiguous or emotionally charged sensory stimuli (e.g., a person smiling) and assess whether faulty mappings (e.g., “smile = threat”) become entrenched. Key metrics include prediction error dynamics, inference accuracy, and persistence of incorrect beliefs across varying contexts will help asses whether the model captures delusional-like behavior.

This project contributes to the field of computational psychiatry, offering an account of how delusions might emerge from interactions between learned mappings, belief uncertainty, and precision-weighted inference. It also bridges insights from clinical psychology, neuroscience, and artificial intelligence, demonstrating the utility of hybrid predictive coding not only as a model of brain function but also as a potential tool for understanding psychopathology. While not claiming to fully replicate the complexity of psychotic experience, the thesis aims to advance the mechanistic understanding of delusions and promote an interdisciplinary dialogue on their origins.

References

[1] J. N. Harding, N. Wolpe, S. P. Brugger, V. Navarro, C. Teufel, and P. C. Fletcher, “A new predictive coding model for a more comprehensive account of delusions,” The Lancet Psychiatry, vol. 11, no. 4, pp. 295–302, 2024. doi: 10.1016/S2215-0366(23)00411-X.

[2] B. Millidge, A. Seth, and C. L. Buckley, “Predictive Coding: A Theoretical and Experimental Review,” 2022. [Online] Available: https://arxiv.org/abs/2107.12979.

[3] A. Tschantz, B. Millidge, A. K. Seth, and C. L. Buckley, “Hybrid predictive coding: Inferring, fast and slow. PLOS Computational Biology, vol. 19, no. 8, e1011280, 2023. doi: 10.1371/journal.pcbi.1011280.