Contribution
The paper names a methodological error and gives it a mechanism. The error is modelling cognition on hand-built representations. Chalmers, French and Hofstadter call the work those representations hide high-level perception: deciding which of the available data are relevant (the “problem of relevance”) and organising them into structure (the “problem of organization”, p. 4–5). They show the error at work in two celebrated programs. Then they argue it cannot be excused as a postponement, because perception cannot be split off from the task that uses it. Finally they present Copycat as an existence proof that the two can be run together. The first two parts are new as a sustained critique; the third summarises work published elsewhere (LIT-435).
Key insight
Representations are the product of perception, and perception is shaped by what the representation is for. A model handed its representation has therefore been handed most of its answer. Analogy-making is the clearest case: the analogy decides which aspects of a situation matter (DNA as a zipper, or DNA as source code, pp. 11, 18), and analogies in turn shape how situations are perceived (“Nicaragua as another Vietnam”, p. 11). So analogy is “not separate from perception: analogy-making itself is a perceptual process” (p. 17).
Assumptions
(Premises of an argument, since the paper proves nothing formally.)
- High-level perception begins “where concepts begin to play an important role” (p. 2). Low-level perception is set aside throughout, and the authors concede a complete model must include it (pp. 2, 30).
- Human high-level perception is flexible: shaped by belief, goals and external context, and capable of radical restructuring (pp. 3–4). The support is examples and older experiments (Bruner’s New Look, Maier’s two-string problem), not new data.
- Representations are short-term, active, working-memory structures, as distinct from long-term knowledge (p. 4); the argument is about the former.
- Fodor’s and Pylyshyn’s encapsulation arguments apply mostly to low-level perception (p. 4). Asserted in one sentence (“Few would dispute”), and load-bearing: the argument needs top-down influence at the conceptual level.
- Microdomains are a legitimate route to the real world (pp. 20–21). The authors argue that “real world” programs like BACON and SME are themselves “stripped-down domains of certain highly idealized logical forms” with English labels attached.
Key results
(What it argues and reports.)
- BACON (pp. 8–10). Its “discovery” of Kepler’s third law used only mean distances and periods, “precisely the data required to derive the law”, plus a bias toward algebraic laws. Kepler took thirteen years among Platonic solids, circles and theology. Qin and Simon’s finding that students given BACON’s data rediscover the laws within an hour is read as a “reductio ad absurdum of the BACON methodology”, not as support for it (p. 10). Similar remarks are said to apply to STAHL and GLAUBER.
- SME (pp. 13–15). In the atom/solar-system example the representations contain almost exactly the relations the analogy needs: “attracts”, “revolves around”, “gravity”, “opposite-sign”, “greater” and “cause” (p. 14). The object/attribute/relation split and SME’s arity matching (a 3-place predicate maps only to a 3-place predicate) force arbitrary choices to which the program is “highly sensitive” (p. 15). The authors credit SME’s mapping work and note its authors made no claims of insight.
- ACME (p. 16) uses a connectionist network for soft-constraint mapping, but on “preordained, frozen structures of predicate logic”.
- Argument 1 against a representation module (p. 17). Perception depends on analogy (the Satanic Verses and Last Temptation of Christ analogy; Saddam Hussein as Hitler or as Robin Hood), so analogy-making cannot be split into “first perception, then mapping”.
- Argument 2 (pp. 18–19). A module would have to supply one context-independent representation. Either it is too narrow for some tasks (zipper vs. source code), or it is all-encompassing, and then selecting from it is “tantamount to high-level perception all over again”. Claimed to apply to “almost any area within artificial intelligence”.
- Copycat (pp. 21–29). Letter-string analogies in a non-circular alphabet. Codelets, chosen nondeterministically from a pool, build bonds, groups, descriptions and correspondences; the Slipnet’s activations spawn and favour codelets (top-down over bottom-up); perception and mapping codelets share the pool; computational temperature, a measure of the amount and coherence of structure, sets the randomness (“parallel terraced scan”, likened to Holland’s two-armed bandit). On “abc → abd, xyz → ?” a snag at z raises temperature, breaker codelets destroy structure, attention floods to Z, and the program can re-perceive xyz as a predecessor group to answer wyz, though “xyd is actually given more often than wyz” (p. 29).
- What Copycat does not do (pp. 29–30). It does not learn; its mechanisms are fixed and hand-coded, which the authors distinguish from BACON’s and SME’s fixed representations. It has no low-level perception; Tabletop, Chapman’s Sonja and Shrager’s laser program move a little toward it.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Most AI models of cognition skip high-level perception by starting from hand-coded representations | moderate | case studies of BACON, SME, ACME; the general claim (“endemic in traditional AI”, p. 20) is asserted from them |
| C2 | BACON’s success at Kepler’s third law is an artefact of pre-selected data and biases, not a model of Kepler’s discovery | strong as a critique of the claim made for it | §2; the authors’ own reading of Qin & Simon (1990) |
| C3 | SME’s representations were built with its target analogies in mind | moderate | Figure 1’s relations; “difficult to avoid the conclusion” (p. 15); no alternative encoding is run |
| C4 | Analogy shapes perception, so perception and mapping cannot be temporally separated | informal argument | anecdotes (pp. 11, 17); no experiment |
| C5 | No context-independent “representation module” can serve all tasks | informal argument, the paper’s strongest | the narrow-or-overloaded dilemma (pp. 18–19) |
| C6 | Copycat builds its own context-dependent representations and integrates perception with mapping | description of an architecture | §4; one worked problem; details cited to Mitchell & Hofstadter (1990) |
| C7 | Copycat’s answer distributions are qualitatively similar to people’s | reported, not shown here | one sentence (p. 28) citing Mitchell (1990) and Hofstadter & Mitchell (1992) |
| C8 | Copycat’s re-perception of xyz is a stripped-down model of a Kuhnian paradigm shift | analogy | p. 29 |
Concepts
- High-level perception. Perception from the level where concepts matter: object recognition, relations, and whole situations; “semantic”, “drawing meaning out of situations” (pp. 2–3).
- Problem of relevance / problem of organization. Which data enter the representation; how they are put into its form (pp. 4–5).
- Meaning barrier. The gap, “rarely crossed by work in AI”, between low-level programs whose representations are not yet meaningful and high-level programs whose meaning is built in (p. 5).
- Representation module. The hypothetical front end that would supply ready-made representations to task processes (p. 7).
- 20–20 hindsight. Representations designed with the answer known.
- Codelet, Slipnet, computational temperature, parallel terraced scan. Copycat’s agents, concept network, randomness control and search strategy (pp. 23–28).
Connections
The paper is the methodological companion to the Copycat paper (LIT-435), which it cites for details; this record holds that paper unread. Its targets are named exactly: Langley, Simon, Bradshaw and Zytkow’s BACON; Falkenhainer, Forbus and Gentner’s SME, built on Gentner’s structure-mapping theory; Holyoak and Thagard’s ACME. It sides with connectionist context-dependent representations (Rumelhart & McClelland, Elman) and Holland’s classifier systems as “a step in the right direction” (p. 7), and cites Marr, James and Lakoff against objectivist representation. Among the formats it lists, frames come from Minsky (LIT-410, filed in the same session); the paper does not cite Minsky, but its target, a representation format whose slots someone else fills, is the frame tradition’s open question. It cites Evans’s ANALOGY from Minsky’s edited Semantic Information Processing as the early analogy program that did build its own representations.
Bearing on the record
- No THEORY document in the record makes a claim this paper supports or contradicts. THEORY-023 is about evidence for machine consciousness, which the paper does not discuss.
- Emergence. Its Copycat section is an engineered case of “higher-level understanding emerges” from local parallel processes with top-down modulation (p. 21), which the record’s emergence surveys (LIT-141, LIT-150) could use as an example; the paper offers no account of emergence beyond the phrase.
- Machine understanding. The “meaning barrier” frames the question that LIT-131 (understanding in deep networks) and Hofstadter’s 2026 essay (LIT-416) ask of learned systems. A THEORY could be drawn, but only with those readings in hand.
- ML practice. It carries an instruction for AI modelling (build
representation-forming into the model; do not hand-code it), addressed to
symbolic cognitive modelling of 1990. No anthology topic holds it well;
the closest is the anthology’s
analysis-and-evaluation, for the charge that a program’s success can be an artefact of its inputs.
Limitations
- Nothing is measured. The Copycat evidence is architectural description and one problem; the human comparison is one sentence citing other work.
- The case studies are few. BACON and SME are typical, the authors say, and ACME and four other programs get a sentence each; “endemic” is inferred from them.
- The Copycat defence is quick. Against the charge that codelets and the Slipnet are themselves hand-coded hindsight, the reply is that mechanisms, unlike representations, are fixed in people too, and their origin is “a question about learning” (pp. 29–30). The Slipnet’s concepts (successor, sameness, leftmost) are the domain’s relevant relations, so the hindsight charge is moved, not answered.
- Argument 2 assumes working memory is small and that selecting from a large representation is as hard as perceiving afresh; both are asserted.
- Preprint only. The published text may differ.
Open questions
- Does Copycat’s architecture scale to domains whose relevant concepts were not chosen by its designers? The paper names Tabletop as a step and leaves it there.
- Where would learned representations from end-to-end training fall in the critique: as the flexible, context-dependent representations it asks for, or as another fixed representation module? The paper’s only evidence on connectionism is its approval of context-sensitivity.
Corrections
- The brief’s citation (authors, journal, 4(3):185–211, DOI) is correct.
- The paper’s reference for Maier (1931) gives the journal as Cognitive Psychology, 12:181–194. It is Journal of Comparative Psychology 12:181–194 (Crossref, DOI 10.1037/h0071361); Cognitive Psychology did not exist in 1931.
- The Kant quotation “Concepts without percepts are empty; percepts without concepts are blind” (p. 31) is a paraphrase. Kant’s sentence (Critique of Pure Reason, A51/B75) is usually rendered “Thoughts without content are empty, intuitions without concepts are blind.”