Predictive Processing

Hierarchical Generative Models

The brain has multi-level models of the world:

High-level (abstract): "There's a dog"
    ↓ predicts
Mid-level (features): "I'll see fur, four legs, tail"
    ↓ predicts
Low-level (sensory): "These pixel patterns"
    ↑
Actual sensory input

Top-down: Higher levels predict lower levels Bottom-up: Prediction errors propagate upward when predictions fail

Precision Weighting

Not all prediction errors matter equally. The brain weights errors by their reliability (precision):

• High precision error → update your predictions (something’s wrong with your model)
• Low precision error → ignore it (probably just noise)

Example: In fog, visual errors have low precision - you trust your prior expectations more than sensory input.

Active Inference

You can reduce prediction error two ways:

1. Update your model (perception/learning)
2. Change the world (action)

Action is prediction error minimization through doing, not just thinking.

Example: Predict “my arm will be up” → move arm up → prediction fulfilled.

Unified Framework

Perception, action, learning, attention - all the same process (prediction error minimization) at different timescales.

Explains Phenomena

• Perceptual inference: Why ambiguous images flip between interpretations
• Attention: Precision weighting determines what you notice
• Learning: Updating generative model based on persistent errors
• Hallucinations: When priors overwhelm sensory input
• Action: Fulfilling proprioceptive predictions

Bayesian Brain

Predictive processing is Bayesian inference:

• Prior beliefs (predictions)
• Likelihood (sensory evidence)
• Posterior (updated belief after seeing data)

Brain does probabilistic inference, not logic.

Perception Studies

• Measure prediction vs. error signals in brain
• Manipulate precision (attention, reliability)
• Test how priors shape perception

Language Processing

• Reading: Predict next word, measure surprise when wrong
• Comprehension: Generate predictions about meaning, update when violated
• Production: Predict articulatory/acoustic outcomes, execute to fulfill

Bilingualism:

• Do L1/L2 have different priors?
• Code-switching as context-dependent precision weighting?
• Error signals stronger in L2 (less precise predictions)?

Psychopathology

• Depression: Overweighted negative priors
• Autism: Prediction-error processing differences
• Psychosis: Prior beliefs too strong (hallucinations) or too weak (delusions)

Computational Modeling

• Formalize as hierarchical Bayesian inference
• Fit models to neural/behavioral data
• Derive predictions about learning rates, attention, perception

Is Everything Prediction?

Does predictive processing really explain all cognition, or just some aspects?

• Maybe perception/action, but does it explain reasoning, planning, memory?

Computational Tractability

Exact Bayesian inference is intractable. How does brain approximate it?

• Various proposals (sampling, variational inference, heuristics)

Content of Predictions

What determines the generative model? Where do priors come from?

• Evolution, development, learning - but this needs explaining

Action Problem

Is action really just prediction fulfillment? What about novel actions, exploration?

• Active inference addresses this but critics say it’s incomplete

Neural Evidence

Is there enough direct evidence that brain implements hierarchical predictive coding?

• Some support, but also alternative interpretations of neural data

Free Energy Principle (Friston)

All life minimizes variational free energy (bound on surprise). Predictive processing follows from thermodynamic imperative to resist entropy.

Very ambitious: Explains life itself, not just cognition.

Predictive Coding

Specific neural implementation: error units, prediction units, hierarchical message passing.

Active Inference

Emphasizes action as prediction fulfillment. Goal-directed behavior emerges from predicting desired states.

Enactivist Predictive Processing (Kirchhoff, Kiverstein)

Combines predictive processing with enactive cognition. Predictions shaped by action-perception loops.