Predictive Processing and Active Inference

Predictive processing models nervous systems as hierarchical generative systems that predict sensory signals and update from error. Perception estimates hidden causes; attention changes the precision assigned to error; action changes sensory input to fulfil predictions. Active inference places perception and control within the same framework.

The framework is powerful because it connects perception, learning, action, and some clinical phenomena. It is philosophically underdetermined. It can be interpreted as literal Bayesian representation, approximate algorithm, dynamical description, or high-level modelling strategy. Its success does not by itself establish internalism, consciousness, or one metaphysics of mind.

Generative models

A generative model represents how hidden causes produce sensory data. The system uses the model in reverse to estimate causes from effects.

Bayes' rule supplies an idealised form:

$$P(h\mid d)\propto P(d\mid h)P(h),$$

where $h$ is a hypothesis about a cause and $d$ sensory data. Priors and likelihoods combine into posterior estimates.

Neural systems need not calculate explicit probabilities. Population activity and dynamics may approximate inference. The claim becomes substantive only when hypotheses, errors, and update rules are independently specified.

Hierarchical prediction error

Higher levels predict activity below; lower levels transmit mismatches. Repeated updating reduces prediction error across a hierarchy of timescales and abstraction.

This architecture explains context effects and perceptual constancy. The same sensory signal is interpreted differently under different priors, while errors correct expectations when evidence is reliable.

A hierarchy can be fitted flexibly after results. Evidence for predictive processing requires characteristic error signals, directional interactions, and intervention on predictions or errors rather than generic top-down influence.

Precision weighting

Precision represents expected reliability of a prediction error. High-precision signals receive greater influence; low-precision signals are discounted. Attention is often modelled as precision control.

Precision offers explanations of selective attention, uncertainty, and context-sensitive updating. It can also become a catch-all: any surprising weighting can be labelled altered precision after the fact.

Quantitative predictions and independent measures are needed to avoid circularity. Precision should not mean merely "whatever made this signal matter."

Active inference

An organism can reduce error by revising its model or by acting so input matches predictions. Eye movements sample expected information; bodily regulation maintains anticipated viable states.

Active inference reframes action as prediction fulfilment under policies. Goals appear as prior preferences over outcomes rather than separate commands.

Critics worry this makes every action self-confirming. Exploration, surprise, and goal change require expected information gain, hierarchical preferences, or learning. The framework needs constraints to distinguish a substantive model from redescribing any control process.

Perception and hallucination

When sensory evidence is appropriately weighted, generative models track the world. When priors dominate or sensory precision falls, internally generated content can become perceptually compelling. This offers models of hallucination and illusion.

The slogan "perception is controlled hallucination" can mislead. Veridical perception is controlled by world-sensitive error and action; hallucination lacks the same external constraint. Shared generative machinery does not erase the epistemic distinction.

The perceptual implications are developed in Embodied, Enactive, and Predictive Perception.

Representation

Predictive models appear representational because they have accuracy conditions and purport to capture hidden causes. Internalists treat them as rich world models.

Enactivists can interpret prediction-error dynamics as organism-environment coordination without contentful inner hypotheses. Instrumentalists treat the formalism as a useful model of behaviour and neural activity.

Evidence that a variable tracks a model parameter does not decide these interpretations. Misrepresentation, decoupled use, consumer mechanisms, and counterfactual intervention strengthen realist content attribution.

Bayesian brain and idealisation

The Bayesian brain thesis ranges from exact probabilistic computation to the modest claim that behaviour approximates Bayesian inference under selected tasks.

Bayesian models can fit many outcomes because priors and likelihoods are flexible. Model comparison should penalise complexity, predict new data, and specify where approximation fails.

Human biases do not straightforwardly refute Bayesian models; bounded systems may use approximations under resource constraints. Conversely, fitting a posterior curve does not show the mechanism represents probabilities.

Consciousness

Predictive processing is not by itself a theory of consciousness. Both conscious and unconscious processing can be prediction-sensitive.

Consciousness has been associated with high-level hypotheses, precision, global error availability, counterfactual depth, or self-models. These are additional proposals that must be compared with workspace, higher-order, and recurrent theories.

The framework may explain content and perceptual presence while leaving phenomenal character open.

Self and emotion

Self-models can be understood as predictions integrating interoception, body ownership, agency, and social information. Mismatches may contribute to altered ownership or depersonalisation.

Emotion can be modelled through interoceptive inference and prediction of bodily needs. Valence may relate to expected regulation, error dynamics, or action tendency.

These models connect domains but risk translating every phenomenon into the same vocabulary. Explanatory value requires domain-specific predictions and mechanisms.

Psychopathology

Predictive accounts have been proposed for psychosis, autism, anxiety, depression, functional symptoms, and many other conditions. Altered priors or precision can often describe symptoms plausibly.

The breadth creates an unfalsifiability concern: opposite symptoms can be attributed to high or low precision at different levels. Clinical heterogeneity and medication effects further complicate inference.

Useful models should identify measurable parameters, predict individual differences, and guide interventions better than descriptive alternatives.

Free-energy formulations

The free-energy principle provides a mathematical framework connecting generative models, inference, action, and self-maintenance. Variational free energy bounds surprise under a model.

Formal generality is not explanatory completeness. A principle compatible with every viable system may constrain models without identifying the particular representations or mechanisms used by a brain.

The distinction between mathematical theorem, modelling assumption, and empirical claim should remain explicit, as in the book's treatment of probability.

Falsifiability

A predictive-processing explanation should state:

• the represented hierarchy or variables;
• prior and likelihood structure;
• precision assignment;
• update dynamics;
• predicted behavioural and neural effects;
• contrasts with a non-predictive alternative.

Without these, "the brain predicted it" can explain any result retrospectively. The framework becomes testable when concrete models risk failure.

Assessment

Predictive processing is best treated as a framework that generates specific models, not a completed theory of mind. Its philosophical value lies in unifying active perception, uncertainty, and control while forcing clarity about the difference between formal description and literal mental representation.

Selected references

• Clark, Andy. Surfing Uncertainty (2016).
• Friston, Karl. "The Free-Energy Principle" (2010).
• Hohwy, Jakob. The Predictive Mind (2013).
• Orlandi, Nico. The Innocent Eye (2014).