paper:friston-2008-free-energy-brain-college-de-franceA Free energy principle for the brain (lecture summary)
TL;DR
Free energy minimization unifies action and perception as two faces of a single optimization principle: an agent suppresses surprise by either updating its internal model (perception) or by acting on the world to sample only expected sensory states (action). Derived from the probabilistic behavior of an ensemble of agents belonging to the same phenotypic class, the free energy bound approximates the log-evidence (marginal likelihood) of a generative model, making it formally equivalent to negative surprise and negative value simultaneously. The lecture series introduces Dynamic Expectation Maximisation (DEM), a variational filtering scheme that inverts nonlinear dynamic causal models in generalised coordinates of motion and yields both time-dependent conditional state densities and time-independent parameter densities, explicitly superseding Kalman and particle filtering for online Bayesian inversion. Presented across three sessions at the Collège de France in May–June 2008, the framework grounds perception in Helmholtz's neural-energy constructs extended through Empirical Bayes and hierarchical generative models, and reframes dopamine not as a reward signal per se but as encoding the conditional precision—certainty—of predictions, consistent with its role in balancing bottom-up sensory drive against top-down empirical priors. This implies that classical and operant conditioning introduce statistical regularities that are learned by the same hierarchical inference machinery used for causal structure, and that rewards are simply predictable (low-surprise) stimuli, making value learning a special case of perceptual inference rather than an independent computational faculty.
What to take away
- 1. Action and perception are mathematically equivalent under a single free-energy principle: both minimize a variational bound on the log-evidence of an agent's generative model of its sensory inputs.
- 2. Free energy, surprise, and negative value are formally identical quantities, meaning maximizing reward and minimizing sensory surprise are the same computation.
- 3. The paper introduces Dynamic Expectation Maximisation (DEM), a variational scheme that performs online Bayesian inversion of nonlinear dynamic causal models in generalised coordinates of motion.
- 4. DEM furnishes time-dependent conditional state densities and time-independent parameter densities simultaneously, a capability Kalman filtering and particle filtering do not provide in a unified online framework.
- 5. The free-energy bound is equivalent to the model's marginal likelihood (log-evidence), directly enabling model selection and Bayesian model averaging without additional approximations.
- 6. Dopamine is hypothesised to encode the conditional precision (certainty) of predictions rather than reward per se, modulating the balance between bottom-up sensory signals and top-down empirical priors during perceptual inference.
- 7. Classical and operant conditioning paradigms are recast as procedures that introduce statistical regularities into the sensorium, learned via the same hierarchical Empirical Bayes machinery used for ordinary causal inference.
- 8. The framework is derived from first principles by considering the probabilistic behaviour of an ensemble of agents sharing the same phenotype, grounding it in population-level thermodynamic reasoning rather than postulated objectives.
- 9. Hierarchical generative models are the substrate the framework requires of the brain: they allow context-sensitive, dynamic construction of prior expectations rather than fixed priors, which is a replicable architectural commitment for computational modelling.
- 10. An open question the lectures raise is whether all neurobiological value and reinforcement substrates can be fully accounted for within the perceptual-inference hierarchy, or whether dedicated value circuitry retains explanatory independence that free-energy minimisation cannot subsume.
Peer brief — for seminar discussion
Friston's 2008 Collège de France lecture summary, spanning three sessions on 29–30 May and 1 June 2008, formalises a single variational principle—free-energy minimisation—that absorbs both perceptual inference and action-selection as special cases. The central claim is that an agent's free energy constitutes a tractable upper bound on the surprise of its sensory exchanges with the environment, and that this bound equals the negative log-evidence of the agent's generative model. Because surprise, free energy, and negative value are numerically identical under this formulation, acting to maximise reward and acting to minimise prediction error are not competing frameworks but the same gradient descent expressed on different variables. The load-bearing technical contribution is Dynamic Expectation Maximisation (DEM), a variational filtering algorithm that inverts nonlinear dynamic causal models by optimising free energy in generalised coordinates of motion, yielding time-dependent conditional state densities and time-independent parameter densities in a single online pass. DEM is explicitly contrasted with Kalman filtering and particle filtering, both of which handle either states or parameters but not both simultaneously in a principled Bayesian way. The hierarchical generative model architecture underpinning DEM instantiates Helmholtz's neural-energy ideas via Empirical Bayes, allowing context-sensitive top-down priors to be constructed dynamically rather than fixed. Two specific mechanistic predictions follow. First, dopamine's functional role is recast: rather than signalling scalar reward prediction error, it encodes the conditional precision—the inverse variance—of predictions, thereby gating the relative weight of bottom-up sensory evidence versus top-down priors. Second, classical and operant conditioning are predicted to be limiting cases of the same hierarchical causal-inference algorithm, with rewards defined simply as statistically predictable (low-surprise) stimuli and aversive events as intrinsically surprising ones. A critical reader would push back on the scope of the precision-weighting dopamine claim: the 2008 summary cites no quantitative neural data or model fits against electrophysiology, making it an interpretive reframing rather than a tested prediction. The framework could in principle have been evaluated here using existing dynamic causal modelling datasets from fMRI or local-field-potential recordings—an alternative the summary does not engage. More broadly, collapsing value and surprise into one scalar sidesteps the question of how the agent specifies which sensory states count as expected (i.e., how phenotypically appropriate priors are set), a degrees-of-freedom problem that the 3-page summary leaves unresolved and which subsequent literature has contested extensively.
Methods (3)
- dynamic expectation maximisation (DEM)A variational approach for dynamic Bayesian inversion of nonlinear causal models, named in this paper.
- hierarchical modelsModels of sensory generation that allow dynamic context-sensitive prior expectations.
- variational filteringMethod to obtain time-dependent conditional densities by maximizing variational free energy.
Frameworks (1)
- Free Energy PrincipleA foundational variational principle from statistical physics that formalizes how self-organizing systems maintain structural integrity and adapt to their environment by minimizing free energy—a mathematical bound on surprise or prediction error. Originally developed by Karl Friston, the framework unifies action, perception, and learning as processes of active inference, where systems both update internal models of the world and act upon it to reduce the divergence between predictions and observations.
Claims (9)
- Perceptual learning is literally an integral part of value learning, necessary to integrate out dependencies on inferred causes of sensory information.
Core unifying claim: perception and value-learning are unified through free energy minimization.
- Acting to maximize value is the same as acting to minimize surprise; value is simply the probability of sensory input expected by an agent.
Reinterprets classical reward/value concepts through free energy lens.
- Exchanges with the environment are maintained within bounds that preserve the integrity of the agent through surprise minimization.
Links free energy minimization to homeostatic preservation of biological systems.
- Acting to optimize value and perception are two aspects of exactly the same principle: minimization of free energy.
Foundational claim unifying action and perception within single optimization framework.
- A system's state and structure encode an implicit and probabilistic model of the environment.
Foundational claim about internal representation emerging from free energy optimization.
- Rewards are simply predictable stimuli (and aversive stimuli are, by definition, surprising)
Redefines reward and punishment in terms of predictability.
- Dopamine encodes the conditional certainty or precision of predictions
Broader role for dopamine beyond reward signalling, influencing top-down/bottom-up balance.
- Free-energy, surprise and negative value are all the same thing
Collapses three quantities into one, emphasizing their equivalence under the principle.
- Value is the probability of sensory input expected by an agent
Redefines value in probabilistic terms, linking to surprise minimisation.
Hypotheses (1)
- Dynamic expectation maximisation can furnish time-dependent conditional densities of system states and time-independent parameter densities through variational free energy optimization in generalised co-ordinates of motion.
Technical hypothesis about DEM method's capacity for online Bayesian inversion.
Original abstract (expand)
Action, perception and free-energy (Thursday May 29th) Value-learning and perceptual learning have been an important focus over the past decade, attracting the concerted attention of experimental psychologists, neurobiologists and the machine learning community. Despite some formal connections; e.g., the role of prediction error in optimising some function of sensory states, both fields have developed their own rhetoric and postulates. In work, we show that perceptual learning is, literally,...
Related work— refs + corpus + external arXiv
Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.
- ≈ 90%
- ≈ 89%
- ≈ 89%
- ≈ 89%
- ≈ 88%
- The Free Energy Principle for Perception and Action: A Deep Learning PerspectiveTim Verbelen, Ozan \c{C}atal, Bart Dhoedt Pietro Mazzaglia2022≈ 88%
- ≈ 88%
- The free energy principle for action and perception: A mathematical reviewChang Sub Kim, Simon McGregor and Anil K. Seth Christopher L. Buckley2017≈ 88%
- ≈ 88%
- Active Inference, Curiosity and Insightin corpus2017≈ 88%
- Active Inference: A Process Theoryin corpus2017≈ 88%
- ≈ 88%
- Kalman filters as the steady-state solution of gradient descent on variational free energyManuel Baltieri and Takuya Isomura2021≈ 88%
- ≈ 87%
- Free Energy Principle for State and Input Estimation of a Quadcopter Flying in WindAjith Anil Meera, Dennis Benders and Martijn Wisse Fred Bos2021≈ 87%
- Reframing the Expected Free Energy: Four Formulations and a UnificationHoward Bowman, Dimitrije Markovi\'c, Marek Grze\'s Th\'eophile Champion2024≈ 87%
- Probabilistic Principles for Biophysics and Neuroscience: Entropy Production, Bayesian Mechanics & the Free-Energy PrincipleLancelot Da Costa2024≈ 87%
- Active Inference and Epistemic Value in Graphical ModelsMagnus Koudahl, Bart van Erp, Bert de Vries Thijs van de Laar2022≈ 87%
- ≈ 84%
- Active inference: demystified and comparedin corpus2021≈ 83%
- ≈ 82%
- Why Learning Requires Feelingin corpus2026≈ 81%
- Collective intelligence: A unifying concept for integrating biology across scales and substratesin corpus2024≈ 81%
- ≈ 81%
- Life as we know itin corpus2013≈ 80%
- Learning without neurons in physical systemsin corpus2022≈ 80%
- Self-Improvising Memory: A Perspective on Memories as Agential, Dynamically Reinterpreting Cognitive Gluein corpus2024≈ 80%
- Emergence and Causality in Complex Systems: A Survey on Causal Emergence and Related Quantitative Studiesin corpus2023≈ 80%
- ≈ 80%
- The computational boundary of a 'self': developmental bioelectricity drives multicellularity and scale-free cognitionin corpus2019≈ 80%
Similar preprints — Semantic Scholar
Cross-corpus bridges (12)
same_concept_as · Nomic cosineExternal markdown files that talk about the same concept as this entity.
- aboutblank_kbFree Energy Principle And Active Inferenceframeworks/free-energy-principle-and-active-inference.md0.881
- aboutblank_kbFree Energy Principleframeworks/free-energy-principle.md0.862
- aboutblank_kbDoes the Free Energy Principle adequately explain morphogenesis and pattern formation in biological systems?questions/does-the-free-energy-principle-adequately-explain-morphogenesis.md0.853
- aboutblank_kbActive Inferenceframeworks/active-inference.md0.850
- aboutblank_kbSurprise Minimization Frameworkframeworks/surprise-minimization-framework.md0.840
- aboutblank_kbFree Energy Minimizationframeworks/free-energy-minimization.md0.836
- aboutblank_kbFree Energy Principleconcepts/cognitive/free-energy-principle.md0.835
- aboutblank_kbFree-Energy Principle Applied To Morphogenesisframeworks/free-energy-principle-applied-to-morphogenesis.md0.821
- aboutblank_kbCan the Free Energy Principle be extended to explain consciousness itself, and what role does predictive processing play in subjective experience?questions/can-the-free-energy-principle-be-extended-to.md0.817
- aboutblank_kbFree-Energy Approach To Pattern Regulationconcepts/systems/free-energy-approach-to-pattern-regulation.md0.804
- aboutblank_kbFree-Energy Approach To Pattern Regulationframeworks/free-energy-approach-to-pattern-regulation.md0.797
- alexanderAn association-based model of dynamic behaviourpapers/extracted/2022-04-05_Stefan-Lesser_tr2011003_abmdb.pdf_993054.md0.753