The Predictive Mind - Hohwy (2013) (EN)

Résumé de lecture, chapitre par chapitre — rapporte ce que le livre dit.

Reading summary, chapter by chapter

Identity

Note: the associated extraction DOSSIER carries a sha256 (bcdb2870eb47c221) and a volume (123,959 words) that differ from those recorded here on the current .txt. The dossier's verbatims are nonetheless present in the file: the 33 quotations retained below have all been re-verified against the current .txt (flattened text).

Table of contents (as read)

Part I — The Mechanism

1. Perception as causal inference
2. Prediction error minimization
3. Prediction error, context, and precision
4. Action and expected experience

Part II — The World

1. Binding is inference
2. Is predicting seeing?
3. Precarious prediction
4. Surprise and misrepresentation

Part III — The Mind

1. Precision, attention, and consciousness
2. Perceptual unity in action
3. The fragile mirror of nature
4. Into the predictive mind

General summary

Hohwy defends a single, deliberately maximalist thesis: the brain is an inference machine, and one mechanism alone — the minimization of prediction error — underlies perception, action, and attention. The starting point is the "problem of perception" cast in cause-and-effect terms: the brain, locked inside the skull, has access only to sensory effects and must reason back to their hidden causes, even though there is no one-to-one relation between causes and effects. The answer is inherited from Helmholtz (via Kant and Ibn al-Haytham): to perceive is to test hypotheses. Through hierarchical internal models, the brain generates top-down predictions of the sensory signal; the gap between prediction and signal — the prediction error — is the only bottom-up signal, and the whole system works to extinguish it.

Part I builds this mechanism piece by piece: from simple Bayesian inference to a hierarchy that learns the world's regularities at different time scales, weighting errors by their precision (the inverse of expected variance), handling context and complexity, and integrating action — active inference — as a second way to reduce error (changing the world to match the model, rather than the model to match the world). Perception and action become "two ways of doing the same thing."

Part II tests the explanatory reach of the mechanism against established debates in cognitive science: the binding problem (binding is inference), cognitive penetrability (to what extent do our expectations shape what we see?), the precarious nature of inference, and misperception. Hohwy shows that the framework opens interesting paths, but insists on fragility: the same mechanism that gives us the world can, when precisions are badly estimated, produce illusions, delusions, and disorders.

Part III turns the gaze inward: attention as precision optimization, unity of consciousness tied to active inference, the indirectness and fragility of our relation to the world, then an extension — explicitly more speculative — toward emotion, introspection, the mental private life, and the self. The book is thus at once an ambitious unifying case and, in places, an exercise in self-limitation: Hohwy repeatedly signals where the demonstration is solid and where it becomes conjectural.

Part I — The Mechanism (ch. 1–4)

Part I motivates, describes, and explains the prediction-error-minimization mechanism, from the simplest perceptual Bayesianism up to a mechanism sensitive to precision, context, and complexity, and integrating action.

• 1. Perception as causal inference — Frames the problem of perception as a problem of causal inference: the brain has only sensory effects and must infer their hidden causes, with no one-to-one relation between the two. Without further constraints, no causal inference is better than any other; the solution is to treat perception as a Bayesian inference constrained by prior expectations, illustrated by binocular rivalry.
• 2. Prediction error minimization — Takes these elements up to formulate the central mechanism: hierarchical perception proceeds by top-down prediction from internal generative models, revision of model parameters, and minimization of prediction error. The hierarchy captures the world's regularities at different time scales through a constant exchange of upward information (error) and downward information (prediction).
• 3. Prediction error, context, and precision — Refines the mechanism: prediction error upper-bounds the "surprisal" the brain cannot evaluate directly; minimizing the error amounts to approaching that bound. It introduces precision (the expected weight of errors according to their variance) and context-dependence, both of which must be learned from the only sensory resources available.
• 4. Action and expected experience — Extends the mechanism to action: error can be reduced not only by changing the model, but by acting to change the sensory input (active inference). Perception and action become two routes of one and the same minimization; the chapter addresses the "dark room" objection (why not simply minimize all surprise by isolating oneself?) and closes Part I on the attractions and challenges of the framework.

Verified quotations

"The problem of perception is a problem because it is not easy to reason from only the known effects back to their hidden causes." (ch. 1)
"In our complex world, there is not a one-one relation between causes and effects, different causes can cause the same kind of effect, and the same cause can cause different kinds of effect." (ch. 1)
"For that matter, a Bayesian approach to perception does not seem to directly concern the full richness of perceptual phenomenology as much as mere conceptual labelling or categorization of causes of input" (ch. 1)
"the central mechanism for hierarchical perceptual inference is prediction on the basis of internal, generative models, revision of model parameters, and minimization of prediction error." (ch. 2)
"The problem of perception begins, as we saw, with the observation that there is not a one-one relation between our sensory input and its causes in the environment." (ch. 2)
"Perceptual inference is a matter of hierarchical prediction error minimization." (ch. 3)
"this learning about context and noise and uncertainty has to be based only on the sensory resources available" (ch. 3)
"Perceiving and acting are but two different ways of doing the same thing." (ch. 4)
"However, we cannot understand how this mechanism works without the idea that the brain does what it does because it needs to minimize surprisal." (ch. 4)
"prediction error bounds surprisal and the only reason for … minimizing the error through perception and action is that this implicitly minimizes surprisal." (ch. 4)

Part II — The World (ch. 5–8)

Part II confronts the mechanism with established debates about our perception of states of the world: binding, cognitive penetrability, the fragility of inference, and misperception. Hohwy insists: a single mechanism suffices, but its correct functioning rests on a delicate balance of precisions.

• 5. Binding is inference — Treats the binding problem as a case of common-cause inference: to bind attributes is to infer that they share one and the same hidden cause, expected precision explaining why binding seems sharp, how it breaks down, and where attention comes in. Binding is described as an inner "fantasy" of statistical origin, shaped by the brain's generative model.
• 6. Is predicting seeing? — Examines cognitive penetrability: if perception depends on prior expectations, do we see what we expect? A nuanced answer — no arbitrary prediction determines perception, and none determines it fully, because prediction error constrains the hypotheses; there nonetheless always remains a slight degree of penetrability so long as uncertainty remains.
• 7. Precarious prediction — Explores the many "knobs" (precision, complexity, hierarchical levels, contexts) that must be tuned to avoid misperception. It applies the framework to disorders — delusion, autism — seen as imbalances in the interplay between sensory input and prior expectations.
• 8. Surprise and misrepresentation — Starts from the observation that we sometimes misperceive: the brain is not an ideal error minimizer. It takes on harder philosophical problems — misrepresentation, rule-following, representation, understanding — notably the case where the error is well minimized yet the percept is nonetheless false (the dog taken for a sheep).

Verified quotations

"The prediction error minimization framework described in Part I is extremely ambitious—it gives the organizing principle for brain function as such." (ch. 5)
"We ended with the idea that binding is essentially a statistically based inner fantasy, shaped by the brain's generative model of the world." (ch. 5)
"So there seems to be some element of binding coming automatically with the perceptual hierarchy, but it is not yet clear what it is or how it connects to the binding problem." (ch. 5)
"If perception depends on prior belief, does it follow that we perceive what we expect to perceive?" (ch. 6)
"Not any old prediction will be able to determine perception, and no prediction can determine it fully." (ch. 6)
"In fact, since there is always some uncertainty in perceptual inference it seems there is always some degree of cognitive penetrability, even if it is only very mild." (ch. 6)
"Perception and action are driven by the simple imperative to minimize prediction error." (ch. 7)
"This chapter explores some of the many different ways we must tend to these various, interacting aspects of prediction error minimization in order to prevent misperception." (ch. 7)
"Prediction error mechanisms are attractive candidates for explaining delusions because delusions seem to concern the delicate interplay between sensory input and the prior expectations we use to make sense of the input." (ch. 7)
"Sometimes we misperceive the world." (ch. 8)
"This tells us that the brain is not a flawless and ideal prediction error minimizer—in relation to some magical brain that knows how sensations are caused." (ch. 8)
"the hypothesis that that thing right there out on the field is a sheep minimizes prediction error right here and now better than the hypothesis that it is a dog, so you end up (mis-)perceiving it as a sheep." (ch. 8)

Part III — The Mind (ch. 9–12)

Part III takes a different perspective from Part II: to turn the gaze inward and apply the framework to deep and recalcitrant aspects of the mind. The framework's maximal explanatory ambition is embraced here, as is a growing admission of speculation.

• 9. Precision, attention, and consciousness — Analyzes attention as the optimization of precision expectations (not as a mental "searchlight" or a bandwidth limit), and its poorly understood relation to consciousness. It draws on the work of Friston and colleagues as a plausible candidate for explaining key findings in the study of attention.
• 10. Perceptual unity in action — Develops the link between error minimization and consciousness, then explains the unity of conscious perception: why the elements hold together (recalling binding) and why there is only a single global perceptual state. Active inference is the key: without action, unity — indeed consciousness — might dissolve.
• 11. The fragile mirror of nature — Brings out the sense in which perception is marked by indirectness and fragility: we are locked inside the skull, with only sensory data to go by, and our reliability rests on the world lending itself to regular statistical inference. It questions our background beliefs about our own body.
• 12. Into the predictive mind — Extends the mechanism, in an acknowledged more "tenuous" way, to emotion, introspection, the mental private life, and the self. Speculative but motivated by the maximal explanatory reach claimed for the framework (in its generalized free-energy formulation).

Verified quotations

"The aim now is to turn the gaze more inwards and explore how the prediction error framework can be of use in understanding deep and puzzling aspects of the mind." (ch. 9)
"This is not a story about bandwidth limits, or some kind of intrinsic salience, or about some kind of mental searchlight taking possession of things." (ch. 9)
"The work of Friston and colleagues suggests that this is a plausible candidate for explaining key findings in the study of attention." (ch. 9)
"This chapter explores the link between prediction error minimization and conscious perception that the previous chapter put on the table." (ch. 10)
"It seems to me Hurley is right that we cannot understand unity unless we factor in the way action and perception work in tandem." (ch. 10)
"it may be that as we descend deeply into more inactive inference it is not just unity but consciousness itself that deserts us—if there is no action then there is no reason for ignition." (ch. 10)
"This reliability belies an underlying fragility borne by the fact we are trapped inside our skulls with only sensory data to go by." (ch. 11)
"noisy messages are the rule in the perceptual life of animals and humans who need to make the best of it." (ch. 11)
"In this last chapter, the prediction error mechanism is extended deep into matters of the mind." (ch. 12)
"These are certainly aspects where the application of prediction error minimization becomes more tenuous." (ch. 12)
"introspection is then only called upon when we suspect that the prediction error arises not in virtue of what the world is like but in virtue of some issue with perceptual inference itself." (ch. 12)

⚠️ What the book does NOT say

It is not a theory of consciousness. Hohwy links prediction error to attention, perceptual unity, and a representationalism about consciousness, but he does not claim to solve the "hard problem." He himself notes that a "very large discussion" of the metaphysics of consciousness stays buried in his remarks, and refers to Nagel, Levine, and Chalmers; the framework explains what conscious content one has, not why there is experience at all.

It is not a complete coverage of the mind. The author explicitly restricts his object: "I focus on key issues in perception but largely leave out higher cognitive phenomena such as thought, imagery, language, social cognition, and decision-making." The Part III extensions (emotion, introspection, the self) are presented as applications where the framework "becomes more tenuous," that is, speculative and unproven.

The book does not centrally rest on the free-energy formalism. Hohwy works mainly with the epistemic version (Bayesian hypothesis testing) and does not in general use the broadened notion of free energy, while affirming: "Fundamentally however there is no difference between these formal frameworks." The book should therefore not be read as a mathematical derivation of the free-energy principle (that is Friston 2010's role); the exposition is deliberately "done with a minimum of technical, formal detail."

It is not a dualism or an idealism of the "brain that invents the world." Despite the emphasis on top-down predictions, Hohwy holds that the world bounds what expectations can impose, and that the mechanism is "just a causal neuronal mechanism" compatible with a reductionist, materialist view. Perception is indirect and fragile, but robustly anchored in the world — not freely constructed.

The book does not take itself as given. Hohwy himself poses the probative question — "What is the evidence that the brain is (only) engaged in prediction error minimization?" — and acknowledges that the burden of proof falls on him: "The burden of the book is then to demonstrate that the prediction error minimization scheme is not merely the best out of a poor bunch of explanations, but is good in its own right." The unifying ambition is stated as a thesis to be defended, not as an established fact.

Provenance and control note

Actual reading depth. Read in full: the Introduction (including the "Plan" and "Background" sections, which set out the architecture and the Helmholtzian genealogy) and the argumentative opening of each of the 12 chapters (statement of object), plus synthesis/conclusion sections of chapters 3, 4, and 5. The internal development, the end-of-chapter notes, the figures, the formal boxes (Bayes' rule, Kullback-Leibler divergence, precision expressions), and the index were only skimmed or consulted spot-wise to locate the quotations. The per-chapter summaries rest on each chapter's opening, on the author's "Plan" section, and on a verbatim extraction dossier; where the internal content was not read in full, the summaries hold to the object announced by the author and fill no gap with invention.

Quotation control. The 33 quotations placed in the "Verified quotations" blocks (10 in Part I, 12 in Part II, 11 in Part III) were re-grepped one by one against the .txt on a flattened copy (whitespace normalized, curly quotes and em/en dashes normalized): 33/33 confirmed (HIT), 0 MISS. The 5 quotations mobilized in "What the book does NOT say" underwent the same control: 5/5 HIT. Total: 38/38 HIT. The "fi"/"fl" ligatures present in the file are kept as-is in the verbatims concerned.

Declared limits. (1) The sha256 (60b501dbd0c0e5fe) and volume (125,791 words) recorded here differ from those of the extraction DOSSIER (bcdb2870eb47c221; 123,959 words) — two states of the file may have coexisted; all quotations were therefore re-verified against the current .txt. (2) No evidence level (E-level) is assigned: this document reports the book's thesis, it does not grade it. (3) Figures, equations, and end-notes are not treated. (4) The internal page numbers cited by the author ("88 The Mechanism," etc.) serve only to locate passages, not as verified page references.