HOW DOES A
LIVING SYSTEM
KEEP ITSELF ALIVE?
The world is uncertain. The organism cannot be.
Temperature changes. Resources disappear. Bodies become hungry. Threats emerge. Sensory information is incomplete.
Yet living systems continuously maintain themselves within a surprisingly narrow range of viable states.
Life exists between uncertainty and stability.
To remain alive, an organism must continuously regulate its internal conditions while interacting with an environment it cannot completely control.
Survival requires more than reacting to the present.
A living system must somehow anticipate what its sensory signals mean, determine what states are likely to follow, and respond in ways that keep the organism within viable bounds.
How can a system remain stable
in a world it cannot predict perfectly?
Karl Friston's Free Energy Principle provides a mathematical framework for approaching this problem.
A SYSTEM CANNOT
ELIMINATE
UNCERTAINTY.
It can become better at living within it.
The Free Energy Principle describes living systems as maintaining themselves by limiting the states they occupy over time.
To do this, the system must continually reconcile what it expects with the sensory evidence it actually encounters.
Living systems maintain their organization by minimizing a quantity called variational free energy.
In FEP, “free energy” is a mathematical quantity used to characterize the relationship between a system's model and its sensory observations.
The system does not have direct access to the world.
It encounters sensory observations and must infer the hidden causes that produced them.
An internal model represents possible causes of sensory input and generates expectations about what the system is likely to encounter.
External causes generate sensory consequences, but those causes are not directly available to the system.
The system must work backward from sensory effects to infer their probable causes.
Sometimes the model gets it wrong.
PREDICTION ERROR
A mismatch provides information that the system's current expectations do not fully explain its sensory observations.
A tractable measure of model–evidence mismatch.
Variational free energy provides an upper bound on sensory surprise, allowing the system to minimize something it can evaluate rather than directly calculating surprise itself.
Minimizing free energy means bringing the system's internal model and its sensory encounters into better statistical alignment.
Same word.
Different concept.
VARIATIONAL
FREE ENERGY
A mathematical quantity related to inference, probability, uncertainty and how well a model accounts for sensory observations.
METABOLIC
ENERGY
Physical energy used by biological systems to maintain cellular activity, neural signaling, movement and other physiological processes.
If sensory evidence is uncertain, how does the system determine what is out there?
It predicts.
THE SYSTEM
DOESN'T JUST
RECEIVE.
It continually generates predictions.
A generative model produces expectations about the sensory signals that hidden causes in the world are likely to generate.
Incoming sensory evidence can then inform whether those expectations adequately explain what the system is encountering.
Prediction is not an event.
It is a loop.
The system continuously coordinates expectations with incoming sensory evidence.
Each pass through the loop changes the conditions under which the next inference occurs. The system is continuously updating rather than solving the world once.
Prediction descends.
Evidence constrains.
TOP-DOWN
PREDICTIONThe generative model supplies expectations about the causes and structure of incoming sensory signals.
BOTTOM-UP
PREDICTION ERRORSensory discrepancies provide evidence that can constrain or revise the system's current expectations.
Perception emerges from the continuing interaction between prior expectations and sensory evidence.
Prediction error
is not simply
failure.
It is information.
A discrepancy indicates that the current model does not completely account for the sensory evidence.
That discrepancy can drive inference and learning, helping the system form expectations that better account for future observations.
The system must estimate what to trust.
Predictive processing therefore depends not only on prediction error, but also on the estimated reliability—or precision—of that error.
LESS RELIABLE
Noisy or uncertain sensory evidence may receive less influence over belief updating.
MORE RELIABLE
Prediction errors estimated to be more reliable can exert greater influence on updating.
In this framework, precision concerns the estimated reliability of a signal or prediction error—often described mathematically in terms related to inverse variance.
What determines
which signals
matter?
Once predictions differ in confidence and sensory errors differ in estimated precision, the system must selectively allocate influence.
This is where the architecture begins to connect directly with questions studied throughout cognitive science.
THIS IS WHERE
COGNITION
ENTERS.
The loop is not cognitively neutral.
Predictions depend on what the system already expects. Sensory evidence differs in reliability. Experience changes future expectations.
These dynamics intersect with core questions in cognitive science: how organisms perceive, attend, learn, remember, form beliefs and select actions.
COGNITION WITHIN A PREDICTIVE SYSTEM
Cognition changes
how the loop
operates.
Rather than treating cognition as a separate layer floating above perception and action, predictive frameworks examine how cognitive processes influence inference throughout the system.
EXPECTATIONS What is already expected?
EVIDENCE What signals arrive?
WEIGHTING How strongly should signals influence inference?
INFERENCE What is currently most probable?
SELECTION What might happen next?
PERCEPTION
Inferring the likely causes of sensory signals.
ATTENTION
Selectively weighting information according to estimated relevance and reliability.
LEARNING
Changing model parameters and expectations through experience.
MEMORY
Past experience contributes to expectations brought into present inference.
BELIEF
Probabilistic expectations about hidden states and their causes.
DECISION
Evaluating possible actions in relation to expected outcomes.
These processes overlap and interact. The diagram does not assign each cognitive function to one isolated stage of inference.
PERCEPTION
Seeing is not simply receiving.
SIGNAL incoming evidence
HYPOTHESIS current best inference
Sensory signals provide consequences of hidden causes, not direct access to those causes themselves.
Prior expectations and sensory evidence jointly constrain what the system currently infers.
Perception reflects the system's current hypothesis about what is most likely generating its signals.
Perception can therefore be examined as inference under uncertainty rather than passive recording of the external world.
ATTENTION
The system cannot treat every signal equally.
Attention can be related to the selective weighting of information: some prediction errors exert greater influence on inference than others.
Within predictive frameworks, the concept of precision provides one way of modeling differences in estimated reliability or confidence.
Multiple sources of sensory information are available.
Signals differ in their estimated reliability and relevance.
Some signals consequently exert greater influence on inference.
Attention can change the effective flow of information through the prediction loop by altering which signals most strongly influence ongoing inference.
The next prediction
carries the
past.
ERROR evidence of mismatch
MODEL future expectations
Stored and learned regularities from prior experience help structure the expectations available to future inference.
BELIEF
×
DECISION
What do I believe is happening— and what next?
In predictive frameworks, beliefs can be understood as probabilistic expectations about hidden states of the world and body.
Those expectations matter because organisms must eventually evaluate and select actions under uncertainty.
This does not mean cognition is reducible to a single prediction-error mechanism. FEP provides a formal systems framework within which these processes can be modeled and related.
Not six separate functions.
One interacting architecture.
From this perspective, cognitive processes operate within a system that continually integrates prior expectations, sensory evidence, uncertainty, learning and action.
What if the system
doesn't change its
belief?
It can act to change the sensory evidence instead.
DON'T JUST
PREDICT
THE WORLD.
ACT ON IT.
Inference does not end with perception.
When expectations and sensory evidence diverge, the organism is not limited to revising its internal model.
It can also act—changing what it samples, where it moves, and the sensory consequences it encounters next.
PERCEPTION AND ACTION BECOME COUPLED PARTS OF AN ONGOING PROCESS OF INFERENCE AND REGULATION.
The system has more than
one way to
respond.
CHANGE
THE MODEL
Sensory evidence can revise beliefs or expectations so the internal model better accounts for what has been observed.
CHANGE
THE SAMPLING
Action can change the organism's relationship with its body or environment, producing different sensory evidence.
These are not mutually exclusive alternatives. Belief updating and action can unfold together within the same recurrent system.
The organism changes
the conditions of its
next inference.
Action alters bodily or environmental conditions, which changes the sensory evidence available to the system.
Action is not simply downstream from cognition. It feeds back into the sensory conditions from which subsequent perception and inference proceed.
Reaching
for a
cup.
Even an ordinary movement changes the sensory conditions available to the organism.
EXPECTED STATE
The system generates predictions about sensory and proprioceptive consequences associated with a possible movement.
MOVEMENT
Motor action changes the body's relationship to the environment and alters subsequent sensory input.
NEW EVIDENCE
Vision, touch and proprioception provide new evidence about the changing state of the movement.
Movement changes sensory evidence. New evidence then informs subsequent inference.
What happens
if I
do this?
The system can evaluate possible courses of action in relation to their expected consequences.
In active inference, a policy refers broadly to a possible sequence of actions the system could pursue. The highlighted policy is a visual example—not an assumed optimal choice.
The system evaluates
possible
futures.
Active inference extends the framework from explaining present observations toward evaluating expected consequences of possible policies.
G(π) represents expected free energy associated with a possible policy π.
MOVE TOWARD
PREFERRED
OUTCOMES.
Policies can be evaluated in relation to outcomes the organism expects to be more consistent with preferred or characteristic states.
REDUCE
UNCERTAINTY.
Policies can also be valuable because they are expected to provide information about uncertain states of the world or body.
Minimizing expected free energy does not mean seeking only familiar or maximally predictable situations. Information-seeking actions can be valuable when they are expected to reduce uncertainty.
Sometimes uncertainty
is exactly where the
system
goes.
Exploration can have epistemic value.
An organism may act specifically to obtain information that resolves uncertainty and improves subsequent inference.
New evidence does not guarantee learning. It provides information that can support subsequent belief or model updating.
Why does any of this
matter for
life?
Physiological states create ongoing regulatory demands.
The system estimates present conditions and possible future states.
Behavior alters the relationship between organism and environment.
Environmental interaction produces new sensory and bodily states.
From this perspective, cognition and behavior participate in the broader regulatory processes through which living systems interact with and respond to changing conditions.
Inference becomes
interaction.
The organism does not merely construct an internal representation of the world. It participates in a continuous exchange in which prediction, perception, action and learning reshape one another.
THE BRAIN DOES
NOT INFER
ALONE.
It predicts through a body, acts through a body, and encounters a world.
HOW DOES A
LIVING SYSTEM
KEEP ITSELF
ALIVE?
We began with uncertainty.
A living organism exists in a changing environment while maintaining the internal conditions necessary for continued viability.
Across this model, we have examined sensing, prediction, inference, learning and action as interacting processes through which an organism responds to that uncertainty.
Brain. Body. Environment.
One coupled system.
The inferential processes described so far do not occur independently of the organism's body or the environment with which it interacts.
ENVIRONMENT
Resources, threats, opportunities, social signals and changing external conditions.
ACTION EXCHANGE
BODY
Interoception, movement, physiology and the regulatory conditions through which the organism remains viable.
REGULATION
GENERATIVE
MODEL
Expectations, inference and model updating contribute to estimates of bodily and environmental states.
Cognition can therefore be examined as participating in an ongoing organism–environment relationship rather than as an isolated computational layer.
The system must
keep some things
within bounds.
Life occupies a limited range of states.
Temperature, hydration, blood chemistry, energy availability and other physiological variables cannot vary without limit.
Organisms continually regulate themselves in relation to these biological constraints.
Variational free energy is not metabolic energy. The Free Energy Principle provides a formal account of inference and self-organization; physiological regulation remains a physical biological process.
Regulation is not
only correction.
It can
anticipate.
HOMEOSTASIS
Regulatory responses can counter deviations after relevant changes have occurred.
ALLOSTASIS
Organisms can also adjust in advance when changing conditions are anticipated.
Predictive regulation provides a useful bridge between inference and biological regulation: organisms need not only respond to present conditions—they can prepare for expected ones.
This connection should not be read as an equation of the Free Energy Principle with homeostasis or allostasis.
COGNITION DOES
NOT OPERATE
IN ISOLATION.
The body changes what can be known and done.
Sensory organs determine what information can be sampled. Bodily states influence ongoing inference. Motor capacities shape the actions available to the organism.
Cognition unfolds within a world.
The environment provides both uncertainty and structure. Action changes what the organism encounters and therefore changes the evidence available for future inference.
From this systems perspective, cognitive activity can be examined as part of an ongoing relationship among neural, bodily and environmental processes.
From uncertainty
to adaptive
interaction.
The model is best understood as a recurrent set of relationships rather than a sequence with a fixed beginning or end.
These are not three independent systems. They describe interacting processes occurring across the same embodied organism–environment relationship.
A powerful framework.
Not the final word.
The Free Energy Principle is unusually broad. That breadth is part of its theoretical appeal and also part of the continuing academic debate surrounding its interpretation and empirical use.
A formal perspective on perception as inference under uncertainty.
A way to relate learning to changes in generative models.
A framework in which action and perception can be modeled together.
A systems perspective linking brain, body and environment.
A formal approach to questions about self-organization and adaptive persistence.
How broadly the principle should be interpreted across cognition and biology.
How specific and testable particular predictions derived from the framework are.
How FEP-based explanations relate to competing cognitive and neuroscientific models.
Which claims follow from the mathematics and which require additional empirical assumptions.
The empirical scope within which particular active-inference models provide useful explanations.
The Free Energy Principle is a theoretical and formal framework. Its breadth should not be taken as evidence that every feature of cognition or biology has been explained by a single mechanism.
COGNITION IS
PART OF A
LIVING SYSTEM.
It unfolds through an organism's ongoing relationship with its body and environment— predicting, sensing, acting, learning and regulating under uncertainty.
The Free Energy Principle offers one way to understand this continuous relationship: a living system attempting to remain viable while navigating a world it can never know with complete certainty.