Simple feature differences such as color, size, motion, and luminance can be detected preattentively, irrespective of the size of the stimulus display. For example, a red x among green x’s is detected as quickly in a large display as in a small display. A number of limitations on preattentive processing are also apparent. First, we are unable to preattentively detect stimuli defined by the absence of a feature (e.g., an O among Qs), or stimuli that are defined by the conjunction of two or more features (e.g., a blue Q among blue and red Os). Do automatic processes behave similarly?
Treisman et al. suggest that preattentive processing is a bottom-up process which reflects the activation of populations of feature detectors. To the extent that automatic and preattentive processes are governed by the same underlying mechanisms, automaticity might reflect the build up of such detectors. If true, the transfer of automaticity to other tasks that reflect preattentive processing would be expected. Treisman et al. investigated this possibility by training subjects in two visual search tasks, and then, after sufficient practice, by seeing whether the targets were detected as rapidly in tasks thought to require preattentive processing. The transfer tasks were a motion detection task and figure ground separation task. Briefly, Treisman et al. found little evidence for transfer. Furthermore, minor changes in the visual search task (e.g., changing the color of the background) disrupted performance. Treisman et al. explain these findings in terms of Logan’s instance theory.
According to feature theory, attention is not required for the detection of simple features, although attention shifts to these features once detected. To illustrate, deciding whether there is a red O among green Os does not require that attention be focused on the red O. Detection is preattentive. By contrast, attention is required for the conjunction of features. For example, attention must be focused on the target stimulus if the task is to locate a blue Q among red Qs and blue and red Os. Memory traces for conjunctive targets should therefore contain more specific information, such as spatial location. Treisman et al. tested this by contrasting visual search for feature and conjunctive targets. In both experiments, two of the four feature targets appeared in two spatial locations on 75% of the trials; the other two targets were equally likely to appear in any of the eight spatial locations.
Treisman et al. (1992) calculated costs and benefits of consistent targets appearing in consistent and inconsistent locations were defined relative to inconsistent targets appearing in random locations. That is,
Consistent location benefit = Inconsistent random – consistent nonrandom
Inconsistent location cost = Inconsistent random – consistent random
The costs and benefits were much greater for conjunctive targets (e.g., a blue Q among red Qs and blue and red Os) than for feature targets (e.g., a blue Q among red Qs). Furthermore, in conjunctive search but not in feature search, there was a small benefit for inconsistent targets appearing in consistent locations when they were the same color as the consistent target, or shared one dimension in common. This suggests that the build-up of position specific traces accounts well for speed up in conjunctive search, but not in feature search. The trace seems to be a conjunction of color, form, and location.
To summarize up to this point, while preattentive and automatic processes are similar in that both occur in the absence of attention, the origin of independence from attention for these processes seems to differ. Consistent with Logan’s instance theory, automaticity, but not preattentive processing, seems to be mediated by the buildup of position-specific memory traces. Treisman et al. (1992) suggest that the origin of independence for preattentive processes is populations of feature detectors.
In conjunctive search, does information irrelevant to the detection decision become part of the memory trace? This question is relevant to the salience of particular features in the memory trace, and was explored in a subsequent experiment. Two targets were used, and one appeared more frequently in two adjacent spatial locations than in the other six locations, while the other target was associated with a some irrelevant feature (e.g., texture). The important point is that the target detection decision could be made without reference to location or to the irrelevant feature. To illustrate, if the target was a pink vertical bar among green vertical bar and pink and green vertical and tilted bars, texture would be irrelevant, even though it was more often broken than filled. Similarly, the detection decision could be made without regard to the target’s location.
Treisman et al. found some benefit for targets that appeared without a particular irrelevant location, or in a high frequency position. That is, subjects responded more rapidly to the target associated with the irrelevant feature when that feature was present, and more rapidly to the target that appeared in the adjacent spatial locations when it appeared in those locations. There was no transfer of benefit associated with the irrelevant feature or frequent location to other targets after the first session. The results from this experiment were contrasted with feature search. Feature targets were associated with a particular conjunction of irrelevant features, but again the presence of these conjunctions was not necessary for the target detection decision. The results can be summarized briefly: there was no benefit associated with the presence of an irrelevant conjunction.
The results of Treisman et al. suggest that automaticity in conjunctive search is mediated by the build up of memory traces that consist of conjunctions of location, color, and dimension. But was conjunctive search fully automatic in the sense that is was spatially parallel? In Experiment 1, display size was varied, and slopes in were reduced substantially—from 100 ms to 20 ms. The non-zero slope nevertheless indicates some attentional involvement. Furthermore, in Experiment 2, the difference between the benefit for consistently located and inconsistently located consistent targets increased with practice. If conjunctive search were to become fully automatic, one might expect equivalent reaction times for consistent targets, irrespective of location.
If conjunctive search remains to some extent serial, then what accounts for the reduced display size slopes found in Experiment 1? Unitization of conjunctions into preattentively available features would results in zero slopes. An alternative possibility, therefore, is that the comparison process becomes more rapid and efficient. By other criteria, search became automatic. For example, scanning for targets shifted from voluntary to involuntary.
What can be concluded about the similarity of automatic and preattentive processes? Automatic processes are highly specific because they do not transfer to other tasks in which the same stimuli are used, and are sensitive to location and dimension characteristics. As Treisman et al. explain, "The effects seem to depend on the formation of new and very specific associations between features, their locations, and the required responses" (p. 360). By contrast, preattentive processes, as in serial search, are relatively insensitive to things like location specificity. A possible implication of these conclusions is that automatic processes require attention for initiation, whereas preattentive processes work, in the truest sense, in the absence of attention.