Controlled Hallucinations and Precision Weighting: Predictive Coding as a Computational Framework for Visual Consciousness

Visual consciousness may arise when the brain's predictive coding system fully minimizes prediction errors, creating a stable generative model state. The brain actively constructs perceptions by comparing top-down predictions with bottom-up sensory inputs, a process sometimes called 'controlled hallucinations.' Evidence from unconscious neural decoding shows predictive processing can occur without conscious vision. Even without attention, the brain automatically generates electrophysiological responses to expectation violations, indicating automatic prediction error generation. Different frequency neural oscillations convey prediction and prediction error signals, supporting intercortical information exchange. Challenges remain, including unclear biological implementation and risk of conceptual circularity. Future research should develop biologically realistic computational models and combine TMS with EEG and fMRI.

Study at a glance

Abstract

This article explores the mechanisms underlying visual consciousness using the predictive coding (PC) theory as a framework. PC theory posits that the brain is a hierarchical hypothesis-testing system that actively constructs visual perceptions, often referred to as "controlled hallucinations," by continuously comparing top-down predictions with bottom-up sensory inputs to generate and minimize prediction errors. Consciousness corresponds to a generative model state in which prediction errors are fully minimized. Existing neuroscience research provides extensive experimental support for this theory. For example, studies have demonstrated, through unconscious neural decoding, that predictive processing can occur even without visual information entering consciousness. Other studies have found that even in the absence of attention, the brain automatically generates early electrophysiological responses associated with expectation violations, suggesting the automatic nature of prediction error generation. Furthermore, the dissociation of neural oscillations of different frequencies in conveying prediction and prediction error signals provides mechanistic evidence for intercortical information exchange. These findings suggest that conscious content corresponds to a generative model state in which prediction errors are minimized. However, this theory still faces significant challenges, including unclear biological implementation mechanisms and the risk of conceptual circularity. Finally, future research needs to develop computational models that are closer to biological reality, combine TMS intervention on specific brain areas, and use Electroencephalography and Functional Magnetic Resonance Imaging recordings simultaneously to deepen the understanding of visual consciousness.