Contribution
It gives the first explicit design for building a perceptual competence out of many small, quasi-independent, specialised processes that run in parallel, each anthropomorphised as a “demon”. None of them has the competence of the whole. It couples that design to learning at three levels: weights, the feature set itself, and the control structure. The feature set is learned by a selection-and-variation process, which is an early statement of evolving the feature detectors and not only their weights.
Key insight
Do not try to specify a pattern in advance; for most human patterns (spoken words, hand-keyed Morse) “the only adequate definition … must be in terms of the consensus of the people who are using it” (p. 514). Instead, build a crowd of cheap feature computers, let the categories vote by weighted shrieks, and let a running score select both the weights and the feature computers. The architecture is chosen for modifiability: an assembly of quasi-independent modules is easier to change than a machine whose parts all interact.
Assumptions
- Task. Supervised pattern classification with a running score, meaning a monitor says when the machine errs (p. 518). Unsupervised operation comes later, and only after supervised training (p. 523).
- Parallelism is assumed natural for data handling and preferable for modifiability (p. 513). Simulation runs serially on an IBM 704 (p. 522 footnote).
- Hill-climbing landscape. In high dimension “the interdependence of the components and the score is so great as to make very unlikely the existence of false peaks completely isolated from the main or true peak”. This is hoped, and is called “a purely experimental question” (p. 521).
- Anthropomorphism is vocabulary. “Useful words to describe our notions” (p. 513).
Key results
- Idealised pandemonium (fig. 1, p. 514). One demon per pattern computes similarity to the image, and the decision demon picks the maximum. Footnote: this is “an exact correlate” of minimum-distance decoding in a signal space (fig. 2). A second footnote rejects “ideal” pattern representatives as “unnecessarily platonic” (p. 516).
- Amended pandemonium (fig. 3, pp. 516–517). Computations the cognitive demons share go to subdemons. There are four levels: data, computational, cognitive and decision. Each cognitive demon’s output is a weighted sum of subdemon outputs, and the weights are its only difference from the others.
- Feature weighting (pp. 517–521). The weight vector over all cognitive demons is searched to maximise the score. Two techniques are given: random sampling (fig. 4), and random sampling followed by small random steps that keep improvements (fig. 5, “a blind man trying to climb a hill”). False peaks (fig. 6), foot-hills and plateaux are discussed. Unicycle versus chess illustrate two kinds of learning difficulty.
- Subdemon selection (pp. 521–522). Each subdemon is given a worth that measures how strongly its output affects decisions (the formula is illegible in the scan). Low-worth subdemons are eliminated. New ones come from “mutated fission”, random changes to a survivor’s parameters after reduction to a canonical form so the program still runs. They also come from “conjugation”, a continuous analogue of one of the ten non-trivial binary functions of two useful subdemons (Table 1).
- Control adaptation (p. 523). Control operations are demons too, and some “will be in a position to change themselves”, which “raises the possibility of irreversible changes”.
- Evolution (p. 523). This is natural selection on processing demons, and could extend to “some crowd” of pandemoniums.
- Unsupervised operation (p. 523). Use the margin by which one cognitive demon “far outshines the rest” as a proxy score.
- Morse pandemonium (fig. 7, pp. 524–525). Dot and dash cognitive demons each take a weighted sum over “some 150 computing subdemons”. Data demons pass mark and space durations. Subdemons are built from a small, “carefully non-binary” vocabulary: degree of equality, of less-than and greater-than, of being the maximum, stored identifications, averages, tracking means, and the last durations identified. Planned extensions add space demons, then about forty character demons under a new decision demon (p. 526).
- Results are given only in the discussion (pp. 530–531). “Improvement does take place; the hill-climbing does work. The sub-demon selection … did not do what we hoped”. The most effective single demon worked “90 per cent of the time”. No data, test set or baseline is given.
- Discussion.
- Bray proposes polynomial regression on durations; Selfridge says “it will not work; we have tried it”, since context extends “at least 10 letters on either side”.
- MacKay proposes “syndicated” learning: start with coupled elements to reduce diversity, then decouple them as competence grows. Selfridge calls it “an extremely accurate and good point”.
- Price distinguishes stochastic from determinate hill-climbing and asks about sample sizes. Selfridge does not answer this directly.
- McCarthy: the demons’ internal work as unconscious thought and their public shouting as conscious thought (p. 527).
- Selfridge says speed “is so completely irrelevant to this problem” and hopes for parallel machines, so that a problem twice as hard needs a machine twice as big (p. 530).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Pandemonium can adaptively improve itself on pattern-recognition problems that cannot be specified in advance | weak (proposal) | design argument; the paper says it “does not … seem on paper to have the same kinds of inherent restrictions”; no results in the paper |
| C2 | An assembly of quasi-independent modules is easier to modify than a densely interacting machine | assertion | stated as a ground for the design (p. 513) |
| C3 | Hill-climbing on feature weights improves performance on hand-keyed Morse | weak | one sentence in the discussion (“the hill-climbing does work”); no numbers beyond the best single subdemon at 90% |
| C4 | Subdemon selection by mutation and conjugation improves the feature set | proposed; reported not to work as hoped | discussion p. 530 |
| C5 | In high-dimensional spaces, isolated false peaks are unlikely | hope | explicitly “a purely experimental question” (p. 521) |
| C6 | The machine can monitor its own performance by the unequivocality of its decisions | proposal | p. 523; conditional on prior supervised training |
| C7 | Demons’ internal processing is unconscious thought; their public shouting is conscious thought | assertion (McCarthy, in discussion) | none; not taken up by Selfridge |
Concepts
- Demon — a quasi-independent process at one level: data, computational (subdemon), cognitive, decision or control.
- Shriek — a cognitive demon’s output, a weighted sum of subdemon outputs.
- Worth — a subdemon’s influence on decisions, used to cull it.
- Mutated fission / conjugation — generating new subdemons by random parameter change, or by a binary logical function of two parents.
- Hill-climbing, foot-hills, false peaks — search vocabulary for feature weighting.
Composition. The whole’s competence, recognition, belongs to no demon. Each cognitive demon’s competence is a weighted vote, and the decision demon’s is taking a maximum. The parts are not minded. The anthropomorphism is declared to be vocabulary, and a subdemon is “reduce[d] … to some canonical form” so that it can be mutated (p. 522). There is a top-level decision demon, so it is a hierarchy with a single arbiter, not a flat society.
Connections
- Minsky (LIT-399, LIT-402, LIT-429). Minsky is thanked first (p. 526). Minsky’s 1977 memo (NOTE-344) has agents “just intelligent enough to accomplish their own specialized purposes”, new agents “by splitting off from old ones, with only small changes”, and perceptron-like detectors “on tap”. These are Selfridge’s subdemons, mutated fission and weighted cognitive demons, without Selfridge’s decision demon on top. The lineage is my pairing; the memo is not read here as citing Selfridge.
- Block (LIT-407). Block’s n. 19 replaces each homunculus with a McCulloch–Pitts “and” neuron. A pandemonium’s demons are already at that level: a weighted sum, a maximum.
- Schwitzgebel (LIT-159, NOTE-131). Not cited. McCarthy’s remark is a broadcast model of consciousness in one sentence. Schwitzgebel (p. 29) treats global-workspace and “fame” models as subsystem-driven, in his reply to Chalmers. McCarthy puts the conscious part in the shouting between demons, which reads more naturally as relational.
- Collective intelligence (LIT-046). A pandemonium is a designed collective whose members are mindless. Selection among a “crowd of” pandemoniums (p. 523) is a second, population level of the same design.
Bearing on the record
- For the bridge. This is the decomposition side’s starting point: competence assembled from parts that lack it, with mental vocabulary explicitly figurative. It is neutral on Schwitzgebel’s C3–C5 because it never has minded parts. Its one consciousness claim (C7) is McCarthy’s aside, and should be cited as that.
- Anthology. It carries machine-learning content: feature weighting by
stochastic hill-climbing, and evolution of feature detectors. The LIT is
tagged
anthology-candidate(see its boundary paragraph).
Limitations
- It is a proposal written in July 1958, before the runs it promises (p. 526). The results are three sentences in the discussion.
- There is no definition of the score beyond “how well the machine is doing the task”, and no data set, error rates or baseline.
- The worth and weight formulas are illegible in the scan read here.
- The no-false-peaks hope in high dimension is unargued, and the author says so.
Open questions
- Did the later runs promised for the November meeting, or later work, show subdemon selection working, or was the feature set fixed in practice?
- Does MacKay’s “syndicated” coupling-then-decoupling change the picture of how parts become independent agents? It is the reverse of fission.
Corrections
- none to a seeded skim (there was no seed)
- Page range. The paper is pp. 513–526. With the title page, the biography and the discussion, the item is pp. 511–531. The citations 511–526 and 511–529 found online are each partly wrong.