Predictive processing and active inference

Research Question

What exactly is the predictive processing / active inference framework, how does it differ from classical feedforward models of perception, and what is the empirical status of the free energy principle as its mathematical foundation?

Findings

(Fill in when completing. Follow the research skill synthesis structure.)

Executive Summary

Predictive processing (PP) is a well-supported computational framework in which the brain maintains hierarchical generative models that continuously predict incoming sensory signals; what propagates upward through the cortex is primarily prediction error (mismatch between prediction and input), not raw sensory data. This is empirically grounded and largely accepted at the level of cortical architecture. Active inference — the proposal that action itself is a form of prediction fulfillment (agents act to make their predictions true, rather than to correct errors passively) — extends this to motor control and decision-making and is supported by growing computational and behavioural evidence. Karl Friston's Free Energy Principle (FEP) is the mathematical superstructure claimed to underpin both: it is mathematically elegant and has been productively applied in computational neuroscience, but faces serious objections about falsifiability and whether its apparent unification is substantive or merely a reframing of existing concepts. The core mechanistic claim (prediction error minimization drives perception) is well-supported; the grand unification claim (all biological self-organization is free-energy minimization) is contested and arguably unfalsifiable in its broadest form.

Key Findings

1. Core PP mechanism (Rao & Ballard 1999) is well-established. The visual cortex implements bidirectional processing: higher areas send down predictions; lower areas send up prediction errors. Feedforward signals carry error (surprise); feedback signals carry predictions. This explains extra-classical receptive field effects and other anomalies that purely feedforward models cannot. The basic architecture is now widely replicated in computational models and consistent with neuroimaging evidence of hierarchical cortical organisation.

2. Active inference reframes action as prediction fulfillment, not error correction. In classical motor control, the brain generates a desired state, detects error, and corrects. In active inference, the brain generates a prediction of a desired future sensory state and then acts to make that prediction come true. Both reduce to the same observable behaviour in simple cases, but active inference generates different predictions in contexts of high uncertainty, exploratory behaviour, and voluntary attention. Experimental evidence for active inference in motor tasks and saccadic eye movements is accumulating.

3. Precision weighting is the PP account of attention. Not all prediction errors are equally weighted: the brain up-weights errors from reliable sources (high-precision signals) and down-weights errors from noisy sources. This maps cleanly onto the neuroscience of attention and provides a PP account of attentional selection without requiring a separate attention module. Aberrant precision weighting has been proposed as a computational mechanism underlying hallucinations (over-reliance on prior predictions, down-weighting of sensory error signals) — directly relevant to Anil Seth's controlled hallucination thesis.

4. Friston's FEP is the mathematical foundation, but its empirical status is contested. The FEP proposes that any self-organising system (brain, cell, organism) necessarily acts to minimise variational free energy — a proxy for surprise. This is mathematically well-defined and has been productively used in computational models. The primary objections are: (a) falsifiability: the FEP can be made to fit almost any system post hoc by appropriately defining the Markov blanket; (b) biological plausibility: the full computation implied by variational inference is not neurally plausible at a detailed circuit level; (c) reframing vs. unifying: critics argue FEP translates reinforcement learning, Bayesian inference, and homeostasis into new vocabulary without resolving their substantive differences.

5. Key empirical support exists, but is predominantly indirect. PP/FEP is supported by: hierarchical predictive representations in visual cortex (fMRI); mismatch negativity (EEG) as a neural correlate of prediction error; perception-as-inference experiments (ambiguous figures, context effects, prior-driven illusions); and active inference models fitting human behaviour in motor and decision tasks. Direct, unique, discriminating empirical tests of the FEP (that no alternative theory can also explain) are rare.

6. PP underdetermines several key questions. The framework is deliberately abstract, and this abstraction has costs: it does not specify which generative models the brain uses, at what level of abstraction, or how the hierarchy is structured. Multiple incompatible implementations of PP could be consistent with the same behavioural data.

Assumptions

• Assumption: The Rao & Ballard (1999) visual cortex model is representative of PP more broadly. Justification: It is the canonical foundational paper and its architecture has been extended and confirmed at multiple scales. However, PP applied beyond visual perception (to action, planning, social cognition) is more speculative and depends on less direct evidence.
• Assumption: The distinction between PP-as-mechanism (prediction error minimization in cortical circuits) and PP-as-grand-theory (FEP as a unified account of all biological self-organisation) is meaningful and useful for evaluating the framework's empirical status. Justification: Critics and proponents both draw this distinction; it is the primary fault line in the debate about the framework's scientific value.

Analysis

Predictive processing as a computational description of cortical processing (Rao & Ballard level) is well-supported and not seriously contested. Feedforward accounts of perception are insufficient; the evidence for hierarchical, bidirectional prediction-and-error processing is robust. This is the safe, empirically-grounded core of PP.

Friston's FEP as a unifying mathematical framework is more ambitious and more controversial. Its mathematical formalism is sophisticated, and it has been productively applied — particularly in computational psychiatry (modelling hallucinations, anxiety, depression as aberrant inference) and in AI/robotics (active inference agents). The falsifiability objection is serious: a framework that can explain any observation post hoc by tuning its parameters is not a scientific theory in the strong sense; it is a mathematical language. This does not make it useless — many useful frameworks in science are not strictly falsifiable at the grand-unification level (general relativity in some regimes, for instance) — but it does mean its epistemic status is closer to "a powerful modelling toolkit" than "an empirically confirmed unified theory."

For Anil Seth's controlled hallucination thesis specifically: PP provides the mechanism (precision weighting, prior-driven inference) that makes the thesis computationally coherent. The thesis is well-grounded given PP; the debate is whether PP's account of consciousness goes beyond a useful metaphor.

Risks, Gaps, and Uncertainties

• The exact neural implementation of variational inference (the computation implied by the FEP) is not established. The FEP is a computational-level description; the implementational-level mechanisms are not well constrained.
• PP provides a framework for explaining existing data, but generating unique predictions (that specifically disconfirm PP if false) is difficult. This limits how rapidly the empirical support base can grow.
• The scope extension of PP to social cognition, language, and consciousness is ongoing and more speculative than the core visual processing application.

Open Questions

• Does the FEP's Markov blanket formalism meaningfully extend to consciousness, or is that a category error?
• What is the specific computational mechanism by which precision weighting is implemented in cortical microcircuits?
• Can active inference produce behaviours that are genuinely distinguishable from classical Bayesian decision theory in real (not simulated) experiments?
• How does PP account for the hierarchical structure of generative models — specifically, what determines the level of abstraction at each cortical tier?

sources

• [x] Rao & Ballard (1999) Predictive coding in the visual cortex
• [x] Clark, A. (2016). Surfing Uncertainty. MIT Press. — consulted via Clark (2013) "Whatever next?" in Behavioral and Brain Sciences (open access)
• [x] Friston, K. (2010) The free-energy principle: a unified brain theory? Nature Reviews Neuroscience
• [ ] Colombo, M. & Series, P. (2012). "Bayes in the brain." British Journal for Philosophy of Science. — not directly consulted; Bayesian brain hypothesis covered via other sources