You are looking at a display of 17 green blobs and one red blob. Your task is to find the red blob and press a key as soon as you have found it. What could be simpler than this visual search task? Its apparent simplicity notwithstanding, this task has opened a fascinating and sometimes complex window into the workings of human visual attention. It is for this reason that we have already discussed visual search multiple times on this blog.
To illustrate why this task is so revealing, now imagine you are looking at a display of 149 green blobs and one red blob. How much longer would it take you to find the red blob now, compared to when there were only 17 green distractors? The figure below shows the two displays:
If your intuition tells you that it would take a roughly equal amount of time to find the red blob irrespective of whether there are 17 or 149 green distractors, then that intuition is correct. This phenomenon is called “popout” and has been known for more than 35 years. Although it may appear intuitively obvious, it has had profound psychological implications. Indeed, on first thought it would seem that the only way the speed of detection can be (nearly) invariant with the number of distractors is if the search proceeds in parallel: Somehow your attentional-visual system must be taking in all the information in each display at once, otherwise the number of distractors must matter.
Now consider a situation in which your task is still to look for a red blob, but the distractors can be any of the colors of the rainbow (or more). What do you think might happen to your search times? Looking at the displays below might give you an idea:
Not surprisingly, under those circumstances people require considerably longer to find the target if there are more distractors. In fact, the additional search time is often a linear function of the number of distractors, such that each additional distractor adds the same increment to total search time. This strong “search slope” regularity has frequently been interpreted as an indication that when a target does not “pop out” because its feature is uniquely different from a uniform background of other features, the search proceeds serially. On this view, each item is examined one-by-one until the target is found (or until all distractors have been ruled out if no target is present).
The issue is not fully settled, however: While there has been little doubt that the popout phenomenon requires a parallel account, it is less clear whether the slower search is truly serial in nature. Specifically, the flat or shallow search slopes with popout searches and the steeper slopes with harder searches can be generated by a parallel process as well, for example by assuming that the attentional resources that can be allocated to each item in parallel decreases with the number of distractors, thereby slowing search.
A recent article in the Psychonomic Bulletin & Review has shed further light on the debate between serial and parallel searches. Researchers Rani Moran and colleagues focused on the diagnosticity of response-time (RT) distributions to differentiate between the different explanations.
The key measure in a visual search task is the time it takes to locate the target (or report its absence). Often, researchers focus on the average of such RTs across many trials of a given type (e.g., with a given number of distractors). But that average is just one limited way of summarizing an entire distribution of RTs, as shown in the figure below:
There are, however, other statistics that can be used to summarize the distribution with greater precision. For example, one can describe the location of its leading edge, the extent of the “tail” with greater-than-average RTs, and how “bulbous” the main part of the distribution is.
Moran and colleagues went one step further and described the entire distribution of RTs using a number of “bins”. The researchers then examined how well a parallel model and a serial model could describe the binned RT distributions in the conditions of a previously-published experiment that is considered a benchmark in the field. In that experiment, participants had to find a “2” among a set of distractors composed of the digit “5”. The figure below shows part of those data:
It can be seen that as the set size increased, search times slowed considerably (and errors, in the bottom panel, increased although that need not concern us here). The data are indicated by black asterisks, and there are 5 observations for each set size: each observation represents one “bin” of the RT distribution. It is clear that the set size effect mainly affects the tail of the distribution (i.e., the top 2-3 points) rather than its leading edge (i.e., the bottom point).
The figure also shows the ability of the two models under consideration to explain the data: The parallel model is represented by red crosses and the serial model (called CGS for “Competitive Guided Search”) by blue diamonds. As a first approximation, both models tend to do well: they both capture the basic set size effect and the fact that this is mainly due to the stretching of the tail of the RT distribution.
Closer inspection, however, reveals that the parallel model fell a little short of accounting for the data: Note how the red crosses are dispersed less than the data for set size 18, but more than the data for set size 3. Overall, if the models are compared quantitatively, so even tiny deviations between data and predictions are added up into the comparison, then the parallel model’s success was significantly poorer than that of the serial model.
Moran and colleagues compared the models’ ability by applying them to several additional classic visual-search results. In all cases, they found that the parallel model performed less well than its serial competitor. Surprisingly, the particular serial model used here—the CGS—outperformed the parallel model even for the pop-out task! These results contradict the intuition for above that the constant search time (when the number of items increases) in the pop-out search task can only be explained by parallel attentional processing. As the authors explain, a serial attention model can account for the constant search time if there is attentional “guidance” towards the target—that is, the target is so much more salient than the distractors that attention guides search to it first.
An issue as important as the debate between parallel and serial models of visual search cannot be resolved by a single study. Nonetheless, it is highly informative that a careful quantitative comparison has consistently favored a serial-search model over its parallel counterpart. This is one instance in which a purely verbal interpretation of search slopes has been found to be borne out by quantitative modeling. Given how close the competition was in quantitative terms, however, the work by Moran and colleagues underscores the need for quantitative modeling in cognitive psychology: Some issues ought not to be resolved by verbal theorizing alone, because our human cognitive apparatus frequently delivers incorrect intuitions about the behavior of models.
After all, if Boeing and Airbus don’t design jet liners on the basis of verbal hypotheses alone, and if the human mind is at least as complex as an A380, why would we settle for anything less than a quantitative model?
Article focused on in this post:
Moran, R., Zehetleitner, M., Liesefeld, H. R., Müller, H. J., & Usher, M. (2015). Serial vs. parallel models of attention in visual search: accounting for benchmark RT-distributions. Psychonomic Bulletin & Review. DOI: 10.3758/s13423-015-0978-1