Source
Chalmers, French & Hofstadter (1992), “High-level perception, representation, and analogy”, LIT-414, read in full in NOTE-359 from the 1991 preprint (page numbers are the preprint’s). Burns (1995), a review of Fluid Concepts and Creative Analogies, LIT-457, read in NOTE-375.
What was actually shown
The hidden work (LIT-414 pp. 4–5). Hand-built representations hide high-level perception: deciding which data are relevant (the “problem of relevance”) and organising them into structure (the “problem of organization”).
BACON (pp. 8–10). Its rediscovery 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” (p. 10). NOTE-359 rates this strong as a critique of the claim made for BACON (C2). Burns agrees: finding the data was the discovery (LIT-457).
SME (pp. 13–15). In the atom and solar-system example the representations contain almost exactly the relations the analogy needs: “attracts”, “revolves around”, “gravity”, “opposite-sign”, “greater”, “cause”. Arity matching and the object/attribute/relation split force arbitrary choices to which the program is “highly sensitive” (C3, moderate; no alternative encoding is run).
No representation module (pp. 17–19). First, perception depends on analogy (Saddam Hussein as Hitler or as Robin Hood), so analogy-making cannot be split into “first perception, then mapping”. Second, a module would have to supply one context-independent representation. Either it is too narrow for some tasks (DNA as a zipper or as source code), or it is all-encompassing, and then selecting from it is “tantamount to high-level perception all over again” (C5, informal, the paper’s strongest).
Copycat (pp. 21–29). Offered as an architecture that builds representations and mappings together, by stochastic codelets biased by a concept network. It is described, with one worked problem; the comparison with people is one sentence citing other work (C6–C7).
What could have come out otherwise. The BACON and SME results could have survived re-encoding by someone who did not know the answer. The paper does not try it.
What this does not say
- That Copycat escapes the charge. Its codelets and Slipnet are hand- coded, and the Slipnet’s concepts (successor, sameness, leftmost) are the domain’s relevant relations, so the hindsight charge is moved, not answered (NOTE-359, Limitations). Burns: SME is faulted because it would run the same with water and ice replaced by A and B, yet the same objection to Copycat is dismissed because an observer would work out that SIGMA means successor; “the symbol grounding problem is not so easily solved” (LIT-457, C4).
- That microdomains are innocent. Burns grants the defence that “real-world” programs also restrict their domains, but microdomains “can contain hidden assumptions”, and it is “not clear to what extent the programs … succeed because they avoid representational issues” (LIT-457, C3).
- That “endemic” is shown. BACON and SME are the case studies; ACME and four other programs get a sentence each (C1, moderate).
- Anything about learned representations. The paper names connectionist, context-dependent representations “a step in the right direction” (p. 7) and says no more. Whether a representation learned by training on many tasks is the flexible kind it asks for or another fixed module is open (NOTE-359, Open questions).
- That the module argument is a proof. It assumes that working memory is small and that selecting from a large representation is as hard as perceiving afresh; both are asserted.
Connections
- Frames. Minsky’s frames (LIT-410) are among the formats the paper names. Its charge is not that frames are the wrong format but that a format leaves the filling unexplained (THEORY-462 for the memo’s own account of defaults).
- Perception and cognition. Block’s border between seeing and thinking (THEORY-454) separates them by format. This paper assumes, in one sentence (“Few would dispute”), that Fodor’s and Pylyshyn’s encapsulation arguments apply mostly to low-level perception, and needs top-down influence at the conceptual level.
- Machine understanding. The paper’s “meaning barrier” frames the question LIT-131 and Hofstadter’s 2026 essay (LIT-416) put to learned systems.
- Emergence. Copycat is an engineered case of “higher-level understanding emerges” from local parallel processes (LIT-141, LIT-150); the paper offers no account of emergence beyond the phrase.
- The Copycat paper (LIT-435) and the reviewed book (LIT-428) are filed unread.