Source
Ballard, Hayhoe, Pook and Rao (1997), LIT-1075 (Behavioral and Brain Sciences 20(4), 723–742), read in NOTE-883: section 3 (Figures 9–11, Table 7) and the argument of sections 1–2. The commentaries were not read.
What was actually shown
The account. At about a third of a second, body movements act as pointers: “Deictic representations allow the brain to leave important information out in the world and acquire it just before it is needed in the cognitive program. This avoids the carrying cost of the information.” Working memory holds a few pointers, not contents.
Re-looking (Figure 9b, seven subjects). Coding each block cycle by fixations, M-P-M-D (model, pick up, model again, drop) holds nothing in memory; P-M-D holds colour, M-P-D position, P-D both. M-P-M-D “is the most frequently used by all the subjects, far outweighing the others”: about 43%, against 28%, 11% and 11%, read off the plot. Pick-up-drop cycles came last in a pattern. Subjects averaged 1.3 looks per block and “sometimes made as many as 18 fixations in the model area” where four memorized subpatterns would need four.
Against speed (Table 7). Per block: M-P-M-D 3 s, P-M-D 2.5 s, M-P-D 2.0 s, P-D 1.5 s. “it is surprising that subjects choose minimal memory strategies in view of their temporal cost, especially because they have been instructed to complete the task as quickly as possible, and memorization saves time.”
Both costs move it. With model and copy 70° apart, looks fell from 1.3 to 1.0 per block. With all blocks one colour, so segments could be chunked, they fell to 0.7, and subjects copied subpatterns without looking. The authors use these against two rival readings: that the eyes are merely idle, and that memory is cheap but unreliable, since longer transit gave fewer checks, not more.
What is kept (Figures 10–11, three subjects). A block’s colour was changed during a saccade on about 25% of model fixations. Before pick-up, fixation time rose about 50 ms; after pick-up, when colour was held, 129 ms, and subjects picked a block of the held colour, “in most instances” unaware of the change. “The important finding is that the information retained from prior saccades is determined by what is currently relevant for the task.”
What could have come out otherwise. People could have memorized subpatterns, the faster route, and looks could have been insensitive to the cost of looking or of holding. Neither happened.
What this does not say
- That memory’s carrying cost is measured. It is read from choices. No utility, fitted model or statistics appear in the article; the main data are seven subjects with simulated blocks, the colour-change data three. The authors cite the same pattern with real blocks in up to 20 subjects (Pelz 1995), not reported here.
- That people minimize memory regardless of time. Gray et al. (THEORY-346) argue the per-block timings here are confounded, since encoding may be amortized over a whole trial; whether re-looking costs more total time is not settled by this design.
- That this is amortization. What stays in the world is raw input, and the computation on it is done anew each time. The article argues against precomputation where needs vary: “An obvious reason for timely, task-dependent computation is that its products are so varied that the brain cannot precompute them.” Its own analogy is demand paging.
- That it generalizes. One task, with random colours chosen to resist grouping; chess and driving are argued by analogy. The neural mapping of section 4 is proposal.
Connections
- QUESTION-095: the “defer and re-read” strategy, the opposite policy to the cockpit’s. It suggests a term the question’s list lacks, how predictable the later need is. Re-reading also defends against staleness: when the world changed, the subject who had not memorized was mostly unaffected.
- THEORY-346: the two accounts differ on whether memory costs anything beyond time; Gray et al. set the soft constraints hypothesis against this one.
- THEORY-360: Kirsh and Maglio change the world to cheapen a computation; here the world is left alone and read again.
Note of 2026-10-20: narrowed and promoted
The promotion audit (batch 063) narrowed the title to what LIT-1075 reports observing, and promoted it on that one study under ADR-047 v3. NOTE-883 read the target article in full, and the review checked Table 7 and §3.2 against the same author-posted PDF. Table 7 is titled “Speed vs. memory tradeoffs observed”, so the per-block times are the authors’ tallies, not a model.
Three clauses of the earlier title left it. “Most often once for colour and once for position” is the authors’ coding of the commonest cycle, not a measurement of what each look fetched. “Keep only the property the current step needs” generalises one property tested in three subjects; the title now gives the measured fixation costs instead. The last was that holding information costs more than re-fetching it, inferred from these choices. That is the reading THEORY-346 contests with a model that prices both in time alone, and the reading Bogacz’s commentary (not read in NOTE-883) puts down to task unfamiliarity. The body keeps it as the authors’ inference. The former promote_when, a powered study with the cost of looking and of holding varied independently and total time measured per person, moves unchanged to QUESTION-235.