Active Inference: The Free Energy Principle in Mind, Brain, and Behavior

Overview

Active Inference: The Free Energy Principle in Mind, Brain, and Behavior by Thomas Parr, Giovanni Pezzulo, and Karl Friston offers a comprehensive and mathematical look at Karl Friston’s theoretical framework for understanding cognition and behavior.

Key Concepts

The Free Energy Principle

• Variational free energy: The principle states that any self-organising system that persists over time must minimise a quantity called variational free energy — an upper bound on surprisal (the negative log-probability of sensory observations given the organism’s generative model). Minimising free energy is equivalent to maximising the evidence for the organism’s model of the world
  • Formal statement: $F=E_q[\ln q(\theta )-\ln p(o,\theta )]\ge -\ln p(o)$ where $q(\theta )$ is the approximate posterior over hidden states, $p(o,\theta )$ is the generative model, and $-\ln p(o)$ is surprisal
  • Biological imperative: Organisms that fail to minimise surprise encounter states incompatible with their continued existence (e.g., a fish out of water); the principle frames survival itself as an inference problem
• Generative models: The brain maintains a probabilistic model of how hidden causes in the world generate sensory data; this model has a hierarchical structure where higher levels encode more abstract, slowly changing regularities and lower levels encode fast sensory fluctuations

Perception and Action as Inference

• Perceptual inference: Perception is the process of updating internal beliefs (the approximate posterior $q$) to better explain incoming sensory data — this is equivalent to minimising free energy by changing the brain’s internal states
  • Prediction error minimisation: Sensory prediction errors propagate up the cortical hierarchy; each level attempts to explain away the errors from the level below by generating top-down predictions
• Active inference: Action is the other way to minimise free energy — instead of changing beliefs to fit sensory data, the organism changes the world (via motor commands) to make sensory data fit its predictions
  • Unification of perception and action: Both perception and action serve the same objective (minimising free energy), abolishing the traditional divide between sensory and motor processing
• Expected free energy and planning: When an agent must choose among future actions or policies, it selects those that minimise expected free energy — a quantity that naturally decomposes into two terms: (1) pragmatic value (achieving preferred outcomes, i.e., reward) and (2) epistemic value (reducing uncertainty about hidden states, i.e., information gain / curiosity)

Attention, Learning, and Precision

• Precision weighting: Not all prediction errors are equally reliable; the brain assigns precision (inverse variance) to different sensory channels — attention is formally cast as the optimisation of precision, amplifying reliable signals and suppressing noisy ones
  • Neurochemical correlates: Neuromodulators like dopamine and acetylcholine are proposed to encode precision; disruptions in precision weighting may explain psychiatric symptoms (e.g., false inferences in psychosis, attenuated sensory precision in autism)
• Learning as model updating: Longer-timescale changes to the parameters and structure of the generative model correspond to learning (synaptic plasticity) and neurodevelopment; the agent refines its model so that future predictions become more accurate
• Hierarchical depth: The depth of the generative model determines the temporal horizon over which an organism can plan and infer; deeper models support more abstract reasoning and longer-range predictions

Applications and Implications

• Computational psychiatry: Active inference provides formal accounts of conditions like schizophrenia (aberrant precision on prediction errors), depression (pessimistic priors about future outcomes), and addiction (biased expected free energy favouring short-term reward)
• Robotics and artificial intelligence: Active inference agents can be built that explore, learn, and act in unfamiliar environments by balancing exploitation and exploration — without needing separate reward functions or curiosity bonuses bolted on
• Biological universality: The authors argue that the framework applies not just to brains but to any system that maintains itself far from thermodynamic equilibrium — from single cells to social systems — suggesting a deep formal continuity across scales of biological organisation

Personal Reflection

[To be added]

Related Books

• Being You - Seth applies predictive processing to consciousness; Active Inference provides the mathematical underpinning
• The Hidden Spring - Solms maps the free energy principle onto brainstem affect — the clinical counterpart to the formal theory
• Everything Is Predictable - Chivers explains the Bayesian reasoning that Active Inference formalises into a theory of mind

---

Parent: Books