Contribution
A contemporary, sympathetic but critical review of LIT-428 by a psychologist who models analogy himself (a postdoctoral fellow in psychology at UCLA, per the reviewer’s note). It gives a compact account of the book’s architecture and thesis, and a diagnosis of why the work had “surprisingly little impact on the field of AI”.
Key insight
Burns’s frame is the Sagan analogy, used both ways: like Sagan, Hofstadter explains ideas to a wide audience and inspires people in his field; like Sagan among some astronomers, he is not taken seriously by many in his field, partly for irrelevant reasons (a Pulitzer Prize and a newsgroup) and partly for valid ones. The valid ones are about method: what a program in a microdomain shows, and how one argues against rival programs.
Assumptions
- That impact on AI is a fair measure to ask of the work, and that intuitions must be argued for if others are to build on them.
- That the same evaluative standard should be applied to one’s own programs as to rivals’. The criticism of Copycat rests on it.
- No assumption about the book is checked against the book here; this note reads the review only.
Key results
- The book (as Burns describes it). Basic Books, New York, 1995, 518 pp., $30, ISBN 0-465-05154-5. A collection of articles, many already published, with “a coherent vision” given by Hofstadter’s prefaces. Represented FARG members: Chalmers, Defays, French, McGraw, Mitchell. Quick tour: chapters 2, 4, 5; chapter 5 (Copycat) “perhaps the best description of an implementation”.
- Parallel terraced scan. Representation starts as unrelated elements and is elaborated by codelets, some of which recognise possible structures and post codelets to build them, others build, and others destroy structures inconsistent with higher-level ones. Codelets wait on a coderack, each with nonzero probability of running; running one changes the probabilities of others. Lowering temperature as coherence grows makes destruction harder. Burns likens it to neural networks and genetic algorithms, and suggests codelets might be learned by genetic algorithms, which would answer the objection that they are hand-coded.
- High-level perception. “Cognition is recognition”: extracting meaning from low-level perception by accessing concepts. Burns traces the idea to the Gestalt psychologists and, with Hofstadter, to Kant. Analogy is not a “heavy weapon wheeled out now and then to deal with especially tough problems” (p. 187, quoted). Slippage between related concepts explains fluidity, and so creativity. BACON is criticised for being given Kepler’s data, when finding the data was the discovery.
- Criticism 1: microdomains. Letter strings, table settings, word puzzles, letter fonts. Hofstadter’s defence, that real-world programs also restrict their domains, less honestly, is granted “a worthwhile point”. But microdomains “can contain hidden assumptions”, and it is “not clear to what extent the programs … succeed because they avoid representational issues that are relegated to low-level perception”.
- Criticism 2: method of argument. Analogy and rhetorical question in
place of analysis of “what a program aims to achieve and how well it
achieves its goals”, sometimes straying “in the direction of ridicule”.
- The Eliza effect (“Preface 4”) is real but “not an all-or-none proposition”; the analysis it calls for should be applied to Hofstadter’s programs too.
- Symbol grounding: SME is criticised because it would run the same with water and ice replaced by A and B; the same objection to Copycat is dismissed because an observer would work out that SIGMA means successor. Burns: “the symbol grounding problem is not so easily solved”. An equally astute observer could infer that the symbol with water and ice means melts.
- Evaluation (“Preface 5”): Hofstadter dismisses models for disagreeing with his intuitions (the index entry “SOAR program … ambitiousness yet boringness of”, p. 515) and says any project is “90 percent” intuition. Burns agrees AI models are hard to compare and validate, but calls the response “tantamount to withdrawal from the argument”.
- Verdict. Recommended; the ideas “might be on the right track”, but “Hofstadter makes his challenge too easy to ignore for those who should take the most notice of it.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The book rests on two ideas, the parallel terraced scan and cognition as high-level perception | moderate | the reviewer’s reading; consistent with the chapter titles in LIT-428 |
| C2 | Hofstadter’s work has had “surprisingly little impact” on AI | assertion | no citation or other measure is given |
| C3 | The microdomain programs may succeed by relegating representation to low-level perception | open question, posed as such | ”It is not clear to what extent” |
| C4 | The symbol-grounding charge against SME applies equally to Copycat | moderate | a symmetric counter-example (SIGMA vs. melts), argued in the review |
| C5 | Hofstadter’s response to the evaluation problem withdraws from argument and risks irrelevance | opinion, argued | the “90 percent intuition” passage and the SOAR index entry |
Concepts
- Parallel terraced scan. Many possibilities explored in parallel at different levels of commitment, with resources shifted by probabilistic choice of codelets.
- Codelet, coderack, temperature. Small specialised pieces of code; the pool they wait in; a global measure of disorder that sets how readily structure is torn down.
- High-level perception. Making sense of perceived material at the level of concepts.
- Eliza effect. Crediting a program with understanding it does not have, because its outputs look meaningful.
- Microdomain. A deliberately small, constructed domain in which a cognitive phenomenon can be modelled.
Connections
- LIT-428 (the book). The work reviewed. The review agrees with LIT-428’s report of it on every point checked; see the LIT’s standing.
- LIT-150 (Rizi, on emergence). LIT-428 pairs the book’s “emergent” with it. Burns’s description adds one thing to check there: the emergence he describes is the absence of “top-level control”, a claim about architecture, not about coarse-graining or prediction.
Bearing on the record
- It does not change LIT-428’s status. The book remains unread, and the review is one reader’s account.
- It sharpens what a reading of LIT-428 should test: whether Copycat’s symbols are grounded in any sense SME’s are not (C4), and whether its success depends on representation decided in advance (C3).
Limitations
- A three-page review: the book’s models are described, not assessed in detail, and no experiment or comparison is reported.
- The impact claim (C2) is unsupported.
- The preface numbering the review uses cannot be checked without the book (see Corrections).
Open questions
- How does the book number its prefaces? Burns’s “Preface 4” (Eliza effect) agrees with LIT-428’s contents note if prefaces take the number of the chapter they precede; his “Preface 5” (evaluation) does not, since the note places that preface before chapter 9.
Corrections
- None to the record. LIT-428’s account of this review was checked against the full text and agrees with it.