Source
Martínez (2025), The information-processing perspective on representation, LIT-149, read in full in NOTE-165 (Philosophy and the Mind Sciences vol. 6).
What was actually shown
The two theses (p. 1). Transmission: representations are primarily signals whose function is to transmit information for its own sake. RDC Trade-Off: they aim at an efficient trade-off of rate (transmission and storage cost), distortion (fidelity, which is what ML calls the loss) and the computational complexity of the coders. The Pareto frontier of the three is the “representational surface” (p. 4).
The criterion. Nonzero mutual information is a by-product of almost any causal connection. Information flow near a rate–distortion optimum, or coders near a complexity–distortion optimum, is “excellent evidence of design or adaptation for the transmission of information for its own sake” (p. 10). The inference is to the best explanation and is not defended against alternatives (NOTE-165, C2).
The one result that is demonstrated (pp. 16–17). By the data-processing inequality, information about a car/no-car decision can only fall from the retina to IT. A mutual-information heuristic would therefore locate the decision at the earliest stage, which is wrong. The paper reads the ventral stream as trading rising distortion for falling decoder complexity until the decision is linearly decodable, which is what neuroscience means by “explicit” (Kriegeskorte & Diedrichsen 2019; Hong et al. 2016).
The mappings (§§4–5, §7). Compression corresponds to cognitive economy (Palmer et al. 2015 on ganglion cells predicting the future; Rosch); channel coding’s “safety bubbles” to representational vehicles (sparse coding, synchronised assemblies); complexity management to explicitness. The classical-versus-4E dispute is recast as a disagreement about the shape of the surface (§6), which footnote 14 concedes is partly terminological.
Teleosemantics (fn. 7, p. 11). Without teleofunctions, “any information processing system can be seen as rate-distortion efficient … with respect to some … distortion function”. Teleofunctions fix the distortion measure.
What could have come out otherwise. Only the data-processing point is a result, and it could not have; it is a theorem. The rest makes no prediction yet, because the complexity leg cannot be computed (fn. 9).
What this does not say
- That RDC adaptations generate the properties of representation. The abstract says so; the body offers three illustrative mappings and leaves others “for another time” (NOTE-165, C8).
- That efficiency alone marks representation. Without teleofunctions it marks nothing, by the paper’s own footnote. The account is a version of teleosemantics, and inherits its open questions.
- What a representation is about, or how it can misrepresent. Neither is addressed.
- Anything about engineered or learned systems. Only evolved systems are treated. Bengio et al. 2013 appears once, in a list “presented without discussion”. Whether a trained autoencoder’s latent code counts is open, and so is whether “linearly probeable” means “represented”; the paper’s footnote 13 cites Ritchie et al. 2020 against that inference.
- That a representational surface has been estimated for any system. Complexity is gauged by heuristics (free parameters, nonlinearity, recurrence), and classical and Kolmogorov complexity are set aside as unusable (fn. 9).
Connections
- Content. Block’s two-factor account (THEORY-089) is the record’s account of what fixes the content of a representation, and it argues that a teleological account cannot be naturalised because selection need not aim mechanisms at truth. That bears on the teleofunctional premise here; the two accounts answer different questions (what a representation is about; what shows a state is one), so no relation is declared.
- Real patterns. Dennett’s compressibility criterion (THEORY-036) is the record’s other compression-based test; it is about which patterns are real, not which states are signals.
- Information bottleneck. NOTE-165 notes the kinship with the information-bottleneck line of deep-learning theory in the anthology (ANTH-LIT-531). The paper does not draw it.
- Frames. NOTE-150 notes that Ivy & Mroczko-Wąsowicz’s frame account (THEORY-593) presupposes some theory of representation and states none; this is one it could use.