Source
Treisman and Gelade (1980), LIT-1082, Cognitive Psychology 12(1), read in full in NOTE-897 from a scan of the typeset article: Experiments I–IV (search), V–VII (texture segregation), VIII–IX (identity and location under brief exposure) and the General Conclusions.
What was actually shown
Feature against conjunction (Experiment I). Distractors were brown Ts and green Xs in near-equal numbers, on cards of 1, 5, 15 or 30 items. A feature target (a blue letter, or an S) gave positive slopes of 3.8 ms/item for colour and 2.5 for shape, not linear in display size. A conjunction target (a green T, sharing its colour with the Xs and its shape with the Ts) gave 28.7 ms/item present and 67.1 absent, linear (over 99% of variance), a ratio of 0.43, which the authors call “quite close to half. These results suggest that search is serial and self-terminating with a scanning rate of about 60 msec per item.” With one item there was no conjunction cost. Up to 13 blocks of practice gave “no sign of a switch from serial to parallel search”.
Difficulty is not seriality (Experiments II–IV). Making a conjunction’s colours hard to tell apart nearly tripled the slopes (55.1/92.4 against 20.5/39.5) but kept them linear with the 1:2 ratio: “we can no longer suppose that search becomes serial only when it is difficult.” A single-feature target hard to discriminate, an ellipse of intermediate size among small and large ones, was slow, up to about 3 s, but negatively accelerated: “search can be slow but need not be serial in the sense of checking each item in turn.” With letters, conjunction sets (distractor parts that could recombine into the target) were slower and linear; sets more confusable with the target by published confusion matrices, but unable to form it, were faster and nonlinear.
Absence. Even in feature search, target-absent trials grew with display size, 25.1 ms/item against 3.1 for present. Misses rose with display size in conjunction search (8.2% at 30 items; 20.7% at 36 for confusable colours).
Conditions. Hand-made cards of stencilled letters in a tachistoscope, at most 30–36 well-spaced items, few and distinct colours, unlimited viewing, accuracy stressed, six subjects per experiment. Seriality is inferred from slopes and their ratio; no eye movements were recorded. Features are detected in parallel only “within the limits set by acuity, discriminability, and lateral interference”. The result could have come out the other way: conjunction slopes could have been nonlinear, could have fallen with practice, or could have tracked confusability rather than conjunction structure. None did.
What this does not say
- That similarity plays no part in search. The paper argues that conjunctions, not similarity, make search item-serial, and its Experiments III–IV are its evidence. The rival account, that difficulty grows continuously with target–distractor similarity and falls with distractor–distractor similarity, with no sharp serial/parallel divide, is Duncan and Humphreys 1989 (LIT-1081). The record could not obtain it and holds the disagreement with one side read. The slopes are the finding; the mechanism, feature maps and serial binding by attention, is Proposed at best on this record’s reading.
- That anything unique in one attribute pops out. Which shape properties are features is, in the authors’ words, not predictable a priori: an S among Ts and Xs popped out, an ellipse differing in size did not. Bricks of one colour that differ in stud count or length may be like the ellipses, not like the S.
- That it holds at the scale of a pile. Nothing was tested beyond 36 items, with ten-way colour discrimination, with similar colours, or with occluding, overlapping items that must be moved.
- That binning converts conjunction search into feature search. The General Conclusions predict, untested, that spatially grouping the distractors would remove the serial scan. It is the authors’ prediction, not their result.
- Anything about deadlines. Viewing was unlimited. The absent-trial slopes show that confirming absence costs a scan; they do not measure search failing under a time limit.
Connections
- CLAIM-310, part 3, rests on this result: sorting on one attribute leaves the wanted item to be told from its binmates by one attribute only. The narrowings above are the ones that claim already states.
- CLAIM-306 takes the absent-trial slopes as the laboratory form of its asymmetry between finding presence and showing absence; CLAIM-307 takes the feature-target result as the case in which retrieval needs no bin.
- QUESTION-114 asks whether a database cost model handles search that pops out. This account is the term such a model would need; the index-selection readings (THEORY-345) have nothing like it.
- QUESTION-113’s retrieval criterion turns on whether an item already stands out. This says when it does, under these conditions, and does not decide between the question’s criteria.