I once asked Anne Treisman, my Ph.D. supervisor, how and when I could get my Ph.D. Anne told me that I will get it once I learn all the good things from her and establish something of my own. This meant that I must learn everything about focused attention then find something new. Considering the status of the feature integration theory in 2000, I knew that my new topic could not be about focused attention. So, I chose to study distributed attention and its implication on scene perception because it is a different mode of attention and relatively less studied compared to focused attention.
Distributed attention is closely related to preattentive processing in Anne’s studies. In describing search asymmetries, Anne coined the term preattentive processing to explain how the visual system pools quantitative values like size in parallel. Although many studies that followed have interpreted preattentive process as a process outside of focused attention (e.g., this paper in AP&P Special Issue), Anne’s preattentive process is not a process outside of focused attention, nor without attention. Unlike Neisser, Anne did not divide visual processing into preattentive and attentive processing dichotomously. In fact, preattentive processing describes a different style of deploying attention, rather than a mechanism separate from attention. One can deploy attention either globally over a scene to extract statistical properties (i.e., preattentive processing) or locally on an object for accurate recognition (i.e., focused attentional processing). People can appreciate beautiful scenery by either perceiving the overall color of the autumn leaves or focusing on the vivid yellowness of a ginkgo leaf.
Preattentive search is search under a distributed attention mode. Specifically, when attention is distributed over a large area of the search display, rather than being narrowly focused and directed serially to one item at a time, the visual system pools feature values of a relevant dimension within a wide attention window. These pooled responses are used to detect the presence of a unique feature. Pooled response including a target is usually larger than that without a target, especially when the target has a sufficiently large feature value from distractors. Observers are thus able to detect the presence of a target by pooling the responses from all items in a display without having to scan each item, producing a relatively flat search slope. You do not need to check every ball to find a basketball among a pool of baseballs, tennis balls, and golf balls. The speed of finding a large basketball does not vary much depending on however many small balls there are in the pool.
One thing to note is that pooled response here is not a summed response, but the average response. If that were the case, the visual system would not be able to distinguish a few optimal values from many less optimal values because they would be similar in terms of a summed response. For example, it would be confusing to distinguish a basketball from many baseballs. Since a large feature value produces a larger response than a small feature value, the summed response of a basketball is similar to that of many baseballs. On the other hand, the average response of a basketball is still larger than that of many baseballs. As such, the average response is more informative and is likely used by the visual system to avoid confusion. Anne suggested that inhibition among feature detectors with similar feature values is a potential physiological mechanism of this averaging.
Recently, Aire Raidvee, Mai Toom, Kristiina Averin, and Jüri Allik found supporting evidence that pooled response to multiple items is an average rather than a sum. They linked Anne’s preattentive processing to Kahneman’s System 1 and investigated Kahneman’s claim that the mean computation, unlike the sum, is effortless. They found that observers were much more accurate in discriminating the mean size of multiple items than the sum of their area or the diameter. This is in line with Anne’s hypothesis that pooled response is an average because it shows that mean computation is relatively easier and less effortful than summation. However, they incorrectly interpreted Anne’s claim about the mean computation by referring it to preattentive processing, which in their claim suggests that mean size can be computed outside of focused attention. Anne’s claim, however, is that mean computation is efficiently achieved when attention is deployed over a large area.
Mean computation can also help people to find a complex target. When you need focused attention to conjoin relevant features and identify a target, you can use the distributed attention mode to scout a large area and facilitate search. When you begin searching for a target, you do not have the vaguest idea where the target is located. In this case, it is better to start by scouting a large area first to identify the potential region of a target, rather than starting to scrutinize from the beginning. Statistical regularities can facilitate the scouting process by summarizing the search display, which may guide our focused attention to the region of the display where the summary suggests a higher probability of the target. This initial scouting maybe even more efficient if the statistical properties of a search display are maintained, which is exactly what Corbett and Munneke have found (although their interpretation was focused on a different topic). They found that finding a target with conjoined features was faster when the statistical property of a search display (i.e., the mean size of a search display) was maintained over trials. Indeed, finding a basketball in your garage after years of living there is a lot easier than finding it right after moving into a new house, because you form a contextual redundancy over many search experiences.
I do not know whether I got my Ph.D. because I fulfilled Anne’s advice on finding my own topic. It seems that the idea of statistical representation (i.e., mean size perception) already incubated in Anne’s idea of preattentive processing. What I have done is to specify the kinds of statistical properties one can extract through this preattentive process. And to better cite the nature of this processing, I realized that it is better to use the term ‘distributed attention’ as opposed to ‘preattentive processing’ to describe the process of summarizing multiple items, because the term ‘preattentive’ carries unintended but misleading connotations such as ‘outside of focused attention’ or ‘without attention’. In conclusion, I think that distributed and focused attention modes are two manifestations of a single attention system that efficiently and flexibly copes with the complexity of a scene.
Psychonomic Society articles focused on this post:
Corbett, J.E., Munneke, J. Statistical stability and set size exert distinct influences on visual search. Attention, Perception & Psychophysics, 82, 832–839 (2020). https://doi.org/10.3758/s13414-019-01905-2
Raidvee, A., Toom, M., Averin, K. & Jüri Allik. (2020). Perception of means, sums, and areas. Attention, Perception & Psychophysics, 82, 865–876. https://doi.org/10.3758/s13414-019-01938-7