Résumé de lecture, chapitre par chapitre — rapporte ce que le livre dit.
THIS DOCUMENT IS NOT A CITATION SOURCE — a reading summary, ungraded.
C'est un résumé de lecture, non coté, qui rapporte ce que le livre dit. Il ne porte aucun poids probant et ne fonde aucune affirmation. Toute citation reste ancrée sur le .txt du livre, jamais sur ce résumé.
Friston (2010) - The Free-Energy Principle - A Unified Brain Theory| Title | The free-energy principle: a unified brain theory? (the title is a question) |
| Author | Karl Friston (Wellcome Trust Centre for Neuroimaging, University College London) |
| Year | 2010 (published online 13 January 2010) |
| Publication | Nature Reviews Neuroscience, vol. 11, Feb. 2010, pp. 127–138 · doi:10.1038/nrn2787 (attested in the file) |
| Type | Review (synthesis article, not an empirical study) |
| Source in corpus | library/39_Predictive_processing/_articles/Friston (2010) - The Free-Energy Principle - A Unified Brain Theory.txt |
| sha256 of the text (16) | dd8d84942cd4d07f |
| Volume | 13,141 words (body ≈ 12,375 excluding references) |
| Structure | Abstract · introduction · ~8 thematic sections · 2 boxes (Box 1, Box 2) · Conclusions · 5 online supplementary boxes |
| Table of contents | ✅ READ (section headers re-read by hand in the .txt) |
Method. Automatic parsing fell back to manual mode: the script tripped over the marginal glossary definitions (Free energy, Homeostasis, Entropy, Surprise…) interspersed within the column. The 8 section headers + 2 Boxes + Conclusions were re-read by hand. The author's "In summary" paragraphs served as anchors for each section. The quotations produced were re-grepped: all confirmed (see control note).
Friston starts from an observation: despite an abundance of data, there are few global theories of the brain. He proposes that the free-energy principle (FEP) is one, and that several major theories — the Bayesian brain, predictive coding, efficient coding (infomax), cell-assembly theory, neural Darwinism, optimal control, reinforcement learning, game theory — converge once one looks at what they optimize: always the same quantity, value (expected reward/utility) or its complement, surprise (prediction error / expected cost).
The foundation is thermodynamic-informational: a biological system must keep its states within physiological bounds, hence occupy a small set of states — a low-entropy distribution. But entropy is average surprise. The brain cannot evaluate surprise directly; yet variational free energy is a computable upper bound on it, borrowed from statistical physics and machine learning. Minimizing free energy therefore minimizes (a bound on) surprise — via two routes: perception (changing the recognition density to predict better) and action (changing the sensory sampling so it conforms to predictions). Crucial point: action can act only on accuracy.
The bulk of the Review then goes through each major theory and shows it to be a special case or a complement of the FEP: the Bayesian brain and predictive coding follow from it (cortical hierarchy = message passing of ascending prediction errors and descending predictions); infomax is a "probabilistic generalization" of it; attention = optimization of expected precision (synaptic gain); neural Darwinism supplies value (the complement of free energy, heritable via genetic/epigenetic priors); optimal control and game theory are rewritten by casting rewards and goals as prior expectations that action is obliged to satisfy (active inference; the simulated mountain car example). The conclusion owns the fact that it is "contrived to highlight commonalities": it is a unification, not a confirmation.
The FEP "accounts for action, perception and learning". The single theme is optimization; the optimized quantity is value, or its complement surprise. Hence the announcement, cautious right down to the title ("?"): these theories might be unified.
"This is the quantity that is optimized under the free-energy principle, which suggests that several global brain theories might be unified within a free-energy framework." (§Abstract)
"In this view, the brain is an inference machine that actively predicts and explains its sensations" (§introduction)
Motivation: resisting a tendency to disorder. A system that persists must keep its sensory states (interoceptive and exteroceptive) within a small number of states — hence at low entropy. Biological systems thus appear to "violate the fluctuation theorem" (a generalization of the second law). Since a system can neither know whether its sensations are surprising nor avoid them directly, the distal imperative (staying within physiological bounds) translates into a proximal imperative: avoiding surprise, moment by moment.
Box 1 / Implications: action and perception. Free energy is an upper bound on surprise; minimizing it makes the recognition density ≈ the Bayesian posterior. Two levers, and only one touches the world: action can reduce free energy only by increasing accuracy — by selectively sampling the predicted data.
"Mathematically, this means that the probability of these (interoceptive and exteroceptive) sensory states must have low entropy" (§Motivation)
"In short, the long-term (distal) imperative — of maintaining states within physiological bounds — translates into a short-term (proximal) avoidance of surprise." (§Motivation)
"The difference between the two densities is always non-negative and free energy is therefore an upper bound on surprise." (§Box 1)
"action can reduce free energy only by increasing accuracy (that is, selectively sampling data that are predicted)." (§Box 1, part b)
"This means that the brain will reconfigure its sensory epithelia to sample inputs that are predicted by the recognition density — in other words, to minimize prediction error." (§Implications)
The brain uses internal generative models to update posterior beliefs in an (approximately) Bayes-optimal way. The free-energy formulation was developed precisely to bypass exact inference by converting it into optimization. A classic objection — that the priors would be arbitrary — "dissolves with hierarchical generative models, in which the priors themselves are optimized". Predictive coding is the implementation: forward connections carrying prediction error, backward connections carrying predictions (Box 2, consistent with the asymmetry of real cortical hierarchies).
"However, this criticism dissolves with hierarchical generative models, in which the priors themselves are optimized" (§Bayesian brain)
"This scheme suggests that the only connections that link levels are forward connections conveying prediction error to state units and reciprocal backward connections that mediate predictions." (§Box 2)
"In short, the free-energy principle entails the Bayesian brain hypothesis and can be implemented by the many schemes considered in this field." (§In summary)
The efficient-coding principle (Barlow's redundancy reduction, formalized as infomax) says that the brain optimizes the mutual information between sensory signals and sparse representations. Friston files it as a special case of the FEP — "a probabilistic generalization of the infomax principle" — obtained when representational uncertainty is ignored and there is no action.
"In this context, the infomax principle becomes a special case of the free-energy principle, which arises when we ignore uncertainty in probabilistic representations (and when there is no action)" (§efficient coding)
"Both are mandated by the free-energy principle, which can be regarded as a probabilistic generalization of the infomax principle." (§In summary)
Cell-assembly theory (Hebb) and associative plasticity are re-read within the FEP: a gradient descent on free energy (changing connections to reduce it) is formally identical to Hebbian plasticity. The invariant regularities of the world (generative-model parameters) must be inferred and encoded in synaptic efficacy. Then attention = optimization of expected precision, implemented as synaptic gain (candidates: neuromodulators such as dopamine, acetylcholine; synchronous activity).
"It transpires that a gradient descent on free energy (that is, changing connections to reduce free energy) is formally identical to Hebbian plasticity" (§cell assembly)
"In summary, the optimization of expected precision in terms of synaptic gain links attention to synaptic gain and synchronization." (§In summary)
Neural Darwinism (selection of neuronal groups, Edelman): a primary epigenetic repertoire, refined by experience-dependent plasticity, selected by re-entrant signalling. As in cell-assembly theory, plasticity rests on correlated pre-/post-synaptic activity, but modulated by value (ascending neuromodulatory systems). Friston bridges the two: free energy is the complement of value, and its long-term average the complement of adaptive fitness. Value/surprise is fixed by the form of the generative model and its priors — heritable.
"This means that free energy is the complement of value, and its long-term average is the complement of adaptive fitness (also known as free fitness in evolutionary biology" (§neural Darwinism)
"value or surprise is determined by the form of an agent's generative model and its implicit priors — these specify the value of sensory states and, crucially, are heritable through genetic and epigenetic mechanisms." (§neural Darwinism)
Optimal control, reinforcement learning (the Bellman equation, temporal difference), game theory and expected utility all start from cost/utility to build value functions that guide action. The FEP reverses the order: it starts from a free-energy bound on the value of states, specified by priors on the motion of hidden states — and these priors can incorporate any cost function. Rewards and goals become prior expectations that action is obliged to satisfy (active inference; the mountain car simulation, FIG. 2).
"More generally, it shows how rewards and goals can be considered as prior expectations that an action is obliged to fulfil" (§optimal control, FIG. 2 legend)
"In this view, the problem of finding sparse rewards in the environment is nature's solution to the problem of how to minimize the entropy (average surprise or free energy) of an agent's states: by ensuring they occupy a small set of attracting (that is, rewarding) states." (§In summary)
Friston owns that the Review is "contrived to highlight commonalities": it shows that many global theories unite under a Helmholtzian perspective (the brain as a generative model of the world). The constant theme: the brain optimizes a (free-energy) bound on surprise, or its complement, value. He opens onto applications (debates over dopamine = reward-prediction error vs surprise; addiction, Parkinson's, schizophrenia) and a programmatic image.
"Although contrived to highlight commonalities, this Review suggests that many global theories of brain function can be united under a Helmholtzian perceptive of the brain as a generative model of the world it inhabits" (§Conclusions)
"it is not difficult to imagine building little free-energy machines that garner and model sensory information (like our children) to maximize the evidence for their own existence." (§Conclusions)
It does not prove the FEP. It is a Review whose title is a question, and whose conclusion calls itself "contrived to highlight commonalities". It unifies theories, it does not test them. "Might be unified", "suggests": the mode is conjectural.
It does not distinguish the FEP's own predictions from those of the sub-theories. On the contrary, Friston presents the FEP as entailing/subsuming the Bayesian brain, infomax, etc. This is exactly the point [[Gershman (2019)]] will raise: showing that a theory subsumes other theories does not give it a distinctive, falsifiable prediction. Do not read this article as an empirical validation of predictive coding.
It does not provide a homogeneous level of empirical evidence. The many results cited (receptive fields, repetition suppression, mismatch, dopamine…) belong to the sub-theories; they do not transfer to the FEP "in general". The empirical credit of a special case is not the credit of the principle.
It is not a neutral source. Friston is the author of the programme; this is an advocacy Review. Its value in the foyer is to state what the founder claims — to be bounded by the critic.
Segmentation: automatic parsing fell back to manual mode (marginal glossary interspersed). The 8 sections + 2 Boxes + Conclusions were re-read by hand in the .txt; the author's "In summary" paragraphs anchored each section.
Anti-fabrication control: all quotations produced were re-grepped against the .txt, on flattened text (two-column review → sentences broken by line breaks and hyphenations free-energy / freeenergy). Confirmed.
Declared limits: the equations, FIG. 2/4 and the 5 online supplementary boxes (S1–S5) are known only through the text. The ~130 references were not treated as content.
Role in foyer: founding source of 39_Predictive_processing, to be read against [[Gershman (2019)]] (boundary kb/39/65). This summary grades nothing.