From Errors to Beliefs and Routes: Rethinking Cortical Inference.

Area of Science:

• Neuroscience
• Cognitive Psychology
• Computational Neuroscience

Background:

• Predictive coding (PC) is a dominant framework explaining perception and cognition through sensory data inference.
• The canonical PC model posits hierarchical exchanges between top-down predictions and bottom-up prediction errors in a stereotyped cortical microcircuit.

Purpose of the Study:

• To critically evaluate the empirical support for the canonical cortical implementation of predictive coding.
• To propose a generalized framework for predictive inference that accommodates diverse neural mechanisms and scales.

Main Methods:

• Review of emerging evidence from visual and auditory systems.
• Analysis of data from learning paradigms and studies on conscious versus non-conscious processing.
• Examination of circuit-level recording and perturbation studies.

Main Results:

• Empirical support for segregated prediction and error units in early sensory cortex is limited and context-dependent.
• Evidence suggests predictive inference is instantiated through multiple neural mechanisms across different scales, circuits, and timescales.
• Current debates about PC often conflate computational, algorithmic, and implementational levels of explanation.

Conclusions:

• The canonical microcircuit model of PC may be too restrictive.
• A generalized predictive coding framework is needed, retaining the inferential core while specifying diverse implementational mechanisms.
• Future research should focus on how predictive influences shape perception, learning, and behavior across various neural substrates and conditions.