Karl Friston
9 min
How can disparate global brain theories, spanning biological, physical, and computational perspectives, be reconciled under a single, unified principle of brain function? The author explores whether a universal free-energy principle can explain the fundamental biological imperative for adaptive systems to resist disorder and maintain homeostasis.
Karl Friston reviews the free-energy principle, which posits that self-organizing systems at equilibrium with their environments must minimize the long-term average of sensory surprise. By deconstructing this principle through three equivalent mathematical formulations—energy minus entropy, surprise plus divergence, and complexity minus accuracy—the author maps it onto existing theories including the Bayesian brain hypothesis, predictive coding, infomax, neural Darwinism, and optimal control theory.
The paper demonstrates that all these major theories share a common focus on optimization. When viewed through the free-energy lens, perception acts to minimize prediction errors (by updating internal states to improve accuracy), while action acts to minimize prediction errors (by changing the sensory input to match predictions). The minimization of free energy provides a formal link between these processes: the brain acts as a Helmholtzian generative model, continuously refining its internal model of the world to maximize sensory evidence. The author shows how this process also subsumes Hebbian learning, attentional gain, and optimal decision-making, framing them as specific ways the brain suppresses free energy.
This principle offers a potential 'grand unified theory' for neuroscience. It suggests that the same mathematical imperative driving basic physiological homeostasis also governs complex cognitive processes, learning, and motor control. By providing a common currency—surprise and its complement, value—it helps researchers identify the shared goals of diverse neuronal mechanisms, potentially offering new insights into neuropsychiatric disorders as failures of optimal inference.
Sam: Its sensory input and an internal probabilistic account of what caused that input. The framework also requires a generative model: an account of how causes produce sensations. The internal estimate works backwards, from sensations towards their likely causes.
Alex: So the model predicts what a cause would produce, and perception asks which cause best explains what arrived.
Sam: Yes, and that gives the framework its two routes. The agent can revise its internal estimates to explain the input better. Or it can act, changing the input to fit its expectations.
Alex: Changing expectations sounds like ordinary learning from evidence. What does changing the evidence look like without becoming wishful thinking?
Sam: Friston’s intuitive example is feeling your way through darkness. You anticipate what you might touch, then move to confirm that expectation. He calls this active inference: acting to sample sensory inputs that match predictions.
Alex: That connects sensing and acting, but it leaves a question. Why would the expectations lead somewhere useful rather than somewhere dangerous?
Sam: In this account, expectations are constrained by the organism’s model. Learning and natural selection shape that model. Some expectations specify states with innate value; others specify how to move or explore until those states are reached.
Alex: Let’s place that against earlier theories. The Bayesian brain already treats perception as estimating causes from evidence. What is added here?
Sam: The Bayesian brain hypothesis says the brain updates probabilistic beliefs using sensory information. Free energy supplies an optimization objective for approximating that inference. The review then places perception and action under that same objective, rather than treating inference alone.
Alex: But a mathematical objective does not tell us what neurons exchange. Does the review offer a proposed implementation?
Sam: It discusses predictive coding, where predictions and prediction errors pass between levels of a brain hierarchy. Higher levels send predictions down. Lower levels send mismatches up, prompting revised predictions that reduce those mismatches.
Alex: So those messages are not interchangeable. One carries an explanation; the other reports what the explanation missed.
Sam: That is the proposed arrangement. But this implementation uses assumptions about the shape of the brain’s probability estimates, specifically a Gaussian approximation. Predictive coding is an implementation considered within the framework, not a direct observation proving the whole principle.
Alex: What concrete evidence does the review use to show that this scheme can actually recover causes?
Sam: One illustration uses a synthetic bird hearing synthetic birdsong. A model generates chirps from underlying control parameters. The listening model then changes its internal estimates to infer those parameters from the sound.
Alex: Because the sounds are generated artificially, there is a known answer. How closely does the listener recover it?
Sam: The example presents three simulated songs. Their underlying causes are identified after around six hundred milliseconds. The reported ninety-percent uncertainty regions encompass the true values used to generate the songs.
Alex: That demonstrates recovery within the simulation. It does not establish that birds use this scheme, or that it outperforms another account.
Sam: That is how I would read it. The comparison is between inferred causes and known generating causes. It is not a biological experiment or a contest between competing brain theories.
Alex: The review also reaches beyond perception. How do learning and attention fit without just renaming everything as inference?
Sam: It distinguishes what gets optimized. Perception updates estimates of changing states. Learning updates the model’s more lasting causal relationships, potentially encoded in connection strengths. In the predictive-coding formulation, correlated presynaptic predictions and postsynaptic errors strengthen connections so predictions suppress errors more effectively.
Alex: That connects with Hebbian plasticity: activity-dependent strengthening of connections. What changes when the question is attention?
Sam: Attention enters through the estimated reliability of prediction errors. More reliable errors have greater influence on the model’s estimates. Friston proposes that changing neuronal responsiveness could implement this weighting, linking the framework with theories of attentional gain.
Alex: And efficient coding already asks the brain to represent sensory information economically. Is that another equivalent theory?
Sam: The connection has conditions. Friston treats efficient coding as a special case when uncertainty is ignored and there is no action. Accurate predictions preserve information, while a penalty for complexity discourages an unnecessarily elaborate model.
Alex: Those conditions seem central to judging the unification. A relationship under particular assumptions is not evidence that all the original theories say identical things.
Sam: The review acknowledges that it is constructed to highlight commonalities. It also says the principle’s capacity to unify perspectives remains to be established. Its strongest contribution is the proposed map of relationships; the unresolved challenge is how those relationships manifest in the brain.
Alex: Given that distinction, who should read the full review, and where should they start?
Sam: Researchers connecting perception, learning, attention, and motor control should read it. Start with “Implications: action and perception,” which explains the alternative routes to reducing free energy. Then read the comparison closest to your field, checking the assumptions that make the connection work.
Alex: And for someone who only needs the central idea, rather than the whole theoretical map?
Sam: Remember the proposal, not a proven universal mechanism: a brain can change its expectations, or act to change what it senses. Friston’s claim is that both serve the same survival-bound objective.
Alex: Keep the connection, and keep its assumptions attached.