Contribution
Dennett’s real patterns were a criterion for the reality of entities the manifest image already posits. This paper turns the criterion into a genealogy: the processes that build the manifest image are themselves pattern extraction. Predictive processing minimises variational free energy, which balances complexity against fit as minimum description length does, so perceiving and learning are compressing. Two philosophical consequences are drawn: a structural, Kantian realism about the manifest image, and a way to see the realist–antirealist dispute about selves as a dispute over criteria. The applications are, in the author’s words, “a prolegomenon to a future cognitive-structuralist metaphysics of the manifest image” (p. 15).
Key insight
If the brain builds its model of the world by compressing sensory data, then what the model posits (doors, cats, a self) are the high-level regularities that make the data compressible. They are real in Dennett’s sense, as gliders are real in the Game of Life. They are real as structure in the data, though, and the data underdetermine what realizes the structure. A self is the common cause that best compresses the sensory effects of one’s own activity. There is such a pattern; whether that makes a self depends on whether one asks for a pattern or a substance.
Assumptions
- A realist, representational reading of predictive processing. Footnote 2 (p. 7) assumes that the theories describe actual neural computation and that the brain builds internal models. It grants that instrumentalist and enactive readings exist, and suggests the view survives on them “in a weaker (perhaps diluted) sense”.
- Optimal codes. The equation of Bayesian inference with description length needs L(x) ≈ −log P(x). Footnote 4 (p. 8) calls whether the brain uses optimal codes “an open question”, and argues that the core idea survives without them.
- The substrate is sensory data. Patterns are patterns in the signal at the sensory periphery, not in the world behind it (p. 11).
- Individualism. Footnote 7 (p. 14) restricts the account to early-developing capacities, object perception and minimal self-modelling, and sets aside the social construction of knowledge, perhaps except for the narrative self.
- Pragmatics fixes the trade-off. Among compressions of one signal the agent’s goals pick the useful one (footnotes 5 and 6, pp. 10–13). Free energy fixes the accuracy–complexity trade-off, but the architecture of the generative model is not derived from it.
Key results
- From K-C complexity to free energy (§2, pp. 4–10).
- Kolmogorov–Chaitin complexity defines a pattern but is not computable, so the paper moves to total description length, TDL(M) = L(M) + L(D|M), and the MDL principle. The toy example: for the string 0110111001101110, model A (repeat 01 eight times) costs 2 + 10, model B (repeat 01101111 twice) costs 8 + 2, and the verbatim string 16 + 0, so MDL prefers the second (pp. 5–6). Checked: the mismatch counts 10 and 2 are right.
- Predictive coding is MDL with prediction error as the data term (pp. 6–7).
- With optimal codes, maximising P(M|D) is minimising −log P(M) − log P(D|M), which is TDL, so “Bayesian inference is real-pattern extraction” (p. 8, footnote 3). Checked: this is the monotonicity of the logarithm, as footnote 3 says.
- Variational free energy splits into a complexity term, KL(Q(M) ‖ P(M)), and an expected negative log-likelihood (p. 10). The first is “analogous to the model length term”, the second to the data term. Checked: this is the standard decomposition. The analogy is the author’s, and the complexity term is a divergence from the prior, not a code length for the model.
- Structural realism about the manifest image (§3, pp. 11–15).
- Generative models “specify probabilistic and causal structural descriptions of sensory patterns” (p. 12). A perceived door is a hypothesis about sensory consequences and a place in a causal network.
- Two senses in which the manifest image is “not freely constructed”: compressibility is a mathematical fact, and the sensory signal is shaped by its external source, “at least unless the model is used to run fully off-line simulations, as may be the case in imagery or dreams” (pp. 12–13).
- Veridicalism, after Chalmers. If two subjects’ sensory streams match, one envatted, they are “necessarily … equally compressible by equivalent higher-level (manifest) structural descriptions”. “the door that your envatted twin opens to enter another room in her virtual world is as real as the door you open” (p. 13).
- What lies outside: what realizes the structural roles (a door or a data structure on a supercomputer), purely intrinsic properties, and deeper structure not induced from the senses, which is noumenal only “in the limited sense” (p. 14).
- Footnote 9 (p. 15): Dennett (2013) allows some manifest properties to be “projections”. Spatial location sits nearer the descriptive end and a baby’s cuteness nearer the projective.
- The self (§4.1, pp. 15–17).
- Some sensory regularities are endogenous: tactile input from typing, vision from one’s own eye movements, interoceptive hunger. “the self is inferred and represented as a common, single cause that underlies certain (endogenously generated) regularities in the sensory signal” (p. 16). The model “achieves compression by positing a single, stable entity as the common cause of many endogenous patterns” (p. 16).
- The hierarchy gives a minimal self at fast timescales and a narrative self at slow ones (p. 16).
- The glider swap. Exchange two cells between two gliders and the game runs on unchanged. Which glider is which has “no further fact within the GoL ‘world’” to settle it (p. 17), as Parfit argued for persons.
- The verdict: anti-realist about selves-as-substances (Letheby & Gerrans) and realist about selves-as-patterns (Hohwy & Michael). “The conflict between those positions is apparent, and arises at a meta-level discussion about which criteria of reality are more reasonable to use for this particular domain” (p. 17).
- Footnote 10 (p. 17): the self is partly self-constituted, “a sort of self-fulfilling prophecy”, since active inference makes behaviour conform to the self-model. In some dissociative states the patterns may be “better compressed by model that posits two or more narrative selves residing within one body”.
- Ordinary objects (§4.2, pp. 18–20). Bird’s (2023) criterion is that “the Xs compose an object Y when the Xs are a non-divisible maximal collection of objects such that relevant information about Y compresses corresponding information about the Xs in an efficient albeit possibly lossy way” (pp. 18–19). A cat compresses; “whiskers + crust + dust” does not. Perception parses objects by the same rule, so it is evidentially sensitive to whether ordinary objects exist, against the debunking argument for nihilism and universalism (Korman).
- Caveat (§5, p. 20). Cognition as compression is older than predictive processing (Barlow, Wolff), so the view may generalise beyond it.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Under optimal codes, MAP inference is minimum total description length | strong: a one-line derivation | p. 8, footnote 3 |
| C2 | Free-energy minimisation balances complexity and fit as MDL does | moderate: an analogy between terms, the complexity term being a KL divergence | p. 10 |
| C3 | Learning a generative model is extracting real patterns from sensory data | moderate, conditional on a realist reading of predictive processing and the C1–C2 analogy | §2, footnote 2 |
| C4 | The manifest image is a structural description of sensory data, equally veridical in matched sensory streams, envatted or not | moderate as a conditional: it follows if content is structural; that content is only structural is assumed | §3, pp. 13–14 |
| C5 | The self is represented as a single common cause compressing self-generated regularities | weak to moderate: an interpretation of Hohwy & Michael and Letheby & Gerrans, not shown | §4.1, p. 16 |
| C6 | Realism and anti-realism about selves agree on the facts and differ on the criterion of reality | moderate as a diagnosis | §4.1, p. 17 |
| C7 | Compression-based composition vindicates ordinary objects and makes perception evidence for them | weak: a sketch, relying on Bird 2023 | §4.2 |
Concepts
- Real pattern: a regularity in a substrate that admits a description shorter than the substrate itself, objective because compressibility is a mathematical fact (pp. 4–5).
- Substrate: the base level in which a pattern is found; here the stream of sensory data.
- Total description length: L(M) + L(D|M), the model’s length plus the data’s length given it.
- Manifest image: after Sellars, the ontology of experience and intuitive theories; normativity is set aside (footnote 1).
- Structural veridicalism: the manifest image’s structural claims are true of whatever produces the sensory effects, in a vat or not (after Chalmers).
- Self-as-pattern / self-as-substance: the regularity a self-model compresses, against a persisting entity that grounds numerical identity.
- Minimal and narrative self: self-models at fast and slow timescales of the hierarchy.
Connections
The paper builds on Dennett’s Real Patterns (1991) and his later endorsement of predictive processing, on Grünwald’s MDL, on Friston’s free energy and its description-length reading (Hinton & Zemel; Graves), on Chalmers’s structuralism and veridicalism (2018, 2022), on Hohwy & Michael (2017) and Letheby & Gerrans (2017) for the self, on Parfit for identity, and on Bird (2023) and Petersen (2019) for composition. The record holds Real Patterns (LIT-220, NOTE-190) and Li and Vitányi’s MDL and Bayes condition (LIT-225, NOTE-801). It does not hold Chalmers’s structuralism, Hohwy & Michael, Letheby & Gerrans, or Bird.
Bearing on the record
- CLAIM-355 (real binding, constructed owner). The exchange cited the
paper “for its treatment of the self as a real pattern whose
representation can exaggerate its unity”. The first half is right. The
second is not in the paper. It never says the representation
exaggerates. It says the self-model compresses by “positing a single,
stable entity as the common cause” (p. 16), and that whether this
makes a real self depends on the criterion. Read against CLAIM-355:
- It supports the claim’s structure. The pattern is real; the single substantial owner is not something the pattern supplies (C5, C6). Footnote 10’s self-constitution through active inference is A26’s point that self-representations are part of the organization they represent.
- It cuts against the word “illusion”. On this account the single common cause is a lossy compression, not a false report, so a self-model positing one owner need not misdescribe anything. That is Drayson’s objection as CLAIM-355 states it, and the paper’s “apparent” conflict (p. 17) puts the self’s reality at the choice of criterion. CLAIM-355 survives only where a misdescription can be shown, as it says itself. The paper offers one candidate: dissociative states in which one self-model compresses worse than two (footnote 10). That is where a single-owner model would be not just lossy but wrong.
- Its self is an individual’s, and it says nothing about distributed agents or collectives.
- THEORY-563 (consciousness as a real pattern, Taylor). Same compression criterion, applied to selves and objects, not to consciousness. This paper does not say what makes a pattern a conscious one, any more than Taylor’s does.
- THEORY-036 (the real-pattern criterion admits too much). The paper meets the worry in the way THEORY-036 predicts, by later additions: the agent’s pragmatic trade-off (footnotes 5, 6) and Bird’s maximality and non-divisibility. So it is a case for THEORY-036, not against it.
- THEORY-064, THEORY-691 (the self as model, the self as abstractum). It gives both a mechanism, compression of endogenous signals by a posited common cause, and sides with neither on whether such a self is real. It calls the choice a matter of criteria.
- CLAIM-413 (grounding as a relation). It cuts in an instructive direction. CLAIM-413 calls an envatted brain’s tree experience “ungrounded with respect to the tree”. The paper’s veridicalism says the envatted door is “as real” as yours, because structure in matched sensory streams is the same. Both can be true, since they differ on the referent: the realizer, ungrounded, or the structural role, grounded in the sense of being constrained by an independent source. It also supports “ungrounded ≠ unconstrained”. The signal constrains the model from outside “unless the model is used to run fully off-line simulations, as may be the case in imagery or dreams” (p. 13).
- CLAIM-412 (experience constituted by interpretive organization). Indirect. The paper says the manifest image is constructed by inference, A3’s thesis A, construction, and it says nothing about whether the construction is experience, thesis B. It uses “conscious experience” only to locate the manifest image (p. 3).
- THEORY-095 (Inverted Earth). The veridicalism is a structural internalism about manifest content, the side THEORY-095 leaves open.
- THEORYs warranted. Two are filed: THEORY-872 (the self as a compression pattern, and the dispute as one of criteria), and THEORY-877 (manifest content as structure in sensory data, equally veridical in matched streams).
- No instruction for machine-learning practice; it uses language models only as an example of compression (p. 4).
Limitations
- No data or model. The self and object sections are, in the author’s words, “an initial sketch” (p. 15).
- The bridge to cognition rests on footnote 2’s realist reading of predictive processing and on footnote 4’s open question about optimal codes.
- The VFE–MDL correspondence is an analogy between terms, not an identity.
- Veridicalism holds only if manifest content is wholly structural. The paper assumes this and does not argue it against non-structural content (colour quality, for instance), which it puts in the noumenal.
- The self account assumes the predictive-processing self-model of the cited authors and does not test it.
- “Compression” does not separate the self as a pattern from any other pattern the same agent tracks. Why a self-model is a self, and not one more posit, is not addressed.
Open questions
- When does a single-owner self-model stop being a lossy compression and become a misdescription? Footnote 10’s dissociative case suggests a test: compare the description lengths of one-self and two-self models of the same behaviour.
- Does manifest content have a non-structural part, so that the envatted twin’s door is real in structure but not in quality?
- Does the projective end of footnote 9 (cuteness) include the felt unity of the self? If so, the self-model would be partly projection, which is where an illusion in CLAIM-355’s sense would sit.