The target article claims that the analysis of visual search performance must proceed to larger functional units, moving from items to fixations. This commentary seconds and extends this claim, drawing on earlier studies of active, continuous search. By shedding new light on mechanisms of target detection and distractor processing, these studies provide complementary evidence about task implementation and top-down control in search.
In continuous search tasks participants are exposed to large arrays of items whose scanning requires extended fixation sequences. Item arrays may be arranged in slim columns through which the eye wanders from top to bottom (e.g., Neisser Reference Neisser1963) or rectangular blocks that are scanned line-by-line, as in reading (Prinz Reference Prinz1986). Items may be letters, digits, or other arbitrary elements. Typically, the task requires scanning through the array for a pre-specified target, and the scan terminates upon its detection. A typical scan will thus encounter a large number of distractors before it eventually comes across the searched-for target.
Performance in such tasks seems to call for a simple and intuitive control model, claiming that searching for something requires forming and maintaining a selective attentional set that highlights target-related information. The searcher may, for example, be seen to form a template of the target and maintain it in a primed, pre-activated condition. With this template in mind, she will then scan the search array until she encounters an item that matches it. According to this view, search is controlled by target-related information and target detection is accomplished through target identification. This seemingly natural view was in fact entailed in Neisser's model of search control (Neisser Reference Neisser1967). Based on Selfridge's Pandemonium architecture (Selfridge Reference Selfridge1959), the model claimed that task implementation activates stimulus analyzers for targets, but not for distractors, yielding deep target processing but only shallow distractor processing. As a result, distractors remain unidentified whereas target detection is brought about by target identification.
However, there is now substantial evidence suggesting that this view is fundamentally mistaken. The evidence applies to both target detection and distractor examination (for overviews, see Prinz Reference Prinz and Dornič1977; Reference Prinz1986).
First, target detection does not require deep modes of target processing. Targets are rather detected without being identified. This conclusion is suggested by several converging observations. (1) Pure detection: In tasks requiring parallel search for several targets simultaneously, participants may detect upcoming target locations without/before knowing which target is actually present. Targets can thus be detected and localized before they can be identified. (Neisser Reference Neisser1963; 1967, p. 100). (2) Pseudotarget detection: When a novel distractor appears in the search array, it is often treated as a target. Novel distractors are thus detected without being searched for (Prinz Reference Prinz1979, Exp. 1; Prinz et al. Reference Prinz, Tweer and Feige1974). (3) Detection at distance: Targets are often detected prior to being fixated. They can thus be detected without being the focus of spatial attention (Prinz Reference Prinz, Groner, Menz, Fisher and Monty1983; Prinz & Kehrer Reference Prinz, Kehrer, Groner and Fraisse1982). In sum, these observations suggest that targets are detected by default, not by identification. Target detection seems thus to rely on representational resources other than selectively primed target templates.
Second, distractor examination does not rely on shallow modes of distractor processing. Distractors are rather processed at considerable depth. This conclusion is suggested by a further set of observations: (1) Distractor complexity: Array complexity (as defined by the size of the distractor set) is one of the most powerful drivers of search performance (Prinz Reference Prinz1979). (2) Distractor redundancy: Sequential redundancy of distractor strings and arrays has been shown to drive performance as well. Importantly, distractor redundancy is automatically detected and exploited (Prinz Reference Prinz1979). (3) Distractor transfer: In paradigms requiring switches between search tasks, substantial transfer is obtained when the tasks share the same distractors (Prinz & Ataian Reference Prinz and Ataian1973). Remarkably, this also applies to visually different distractors that share the same names (Prinz et al. Reference Prinz, Hartlich and Lahmeyer1972). In sum, these observations suggest that distractor-related information is not discarded at all. Instead, it seems to be automatically identified within fixations, updated and integrated across fixations, and stored across consecutive trials.
Taken together, these findings suggest a framework for task implementation and top-down control that is actually the inverse of Neisser's intuitive model (Prinz Reference Prinz and Dornič1977; Reference Prinz1986; for overviews of more recent evidence on information integration and task-driven control see Schneider, Reference Schneider2013; Schneider et al. Reference Schneider, Einhäuser and Horstmann2013). This framework builds on three major claims: (1) Incoming information is continuously processed and integrated over space and time (i.e., within and across fixations, respectively). (2) Integration processes generate, maintain and update a dynamic forward model of current and impending distractor information (simultaneously for features, items, strings, etc.). (3) Once such a model is in place subsequent fixation samples are tested against it. As long as they match it, the scan is continued (= pure distractor samples). However, when a sample fails to match it, the scan gets disrupted for closer examination (= mixed distractor/non-distractor samples).
Regarding target processing, this framework accounts for detection by default. Targets are oddballs that fail to match the dynamic distractor model, be it at feature, item, or string level. Regarding distractor processing, it accounts for task implementation and top-down control. Task implementation arises as a by-product of automatic distractor processing: Continuous integration of distractor information within and across fixations generates a dynamic distractor model that instantiates the control structure for the current task. Top-down control is then based on using this model for testing upcoming fixation samples.
These are lessons from continuous search. Why should we take them to heart? On the one hand, continuous search differs from discrete search in several important respects so that only some of these lessons may be directly applicable to that domain. On the other hand, continuous search must be considered a powerful paradigm in its own right. It addresses processing mechanisms for active, extended search as they underlie natural search episodes in real-world scenes and settings.
You have
Access
The target article claims that the analysis of visual search performance must proceed to larger functional units, moving from items to fixations. This commentary seconds and extends this claim, drawing on earlier studies of active, continuous search. By shedding new light on mechanisms of target detection and distractor processing, these studies provide complementary evidence about task implementation and top-down control in search.
In continuous search tasks participants are exposed to large arrays of items whose scanning requires extended fixation sequences. Item arrays may be arranged in slim columns through which the eye wanders from top to bottom (e.g., Neisser Reference Neisser1963) or rectangular blocks that are scanned line-by-line, as in reading (Prinz Reference Prinz1986). Items may be letters, digits, or other arbitrary elements. Typically, the task requires scanning through the array for a pre-specified target, and the scan terminates upon its detection. A typical scan will thus encounter a large number of distractors before it eventually comes across the searched-for target.
Performance in such tasks seems to call for a simple and intuitive control model, claiming that searching for something requires forming and maintaining a selective attentional set that highlights target-related information. The searcher may, for example, be seen to form a template of the target and maintain it in a primed, pre-activated condition. With this template in mind, she will then scan the search array until she encounters an item that matches it. According to this view, search is controlled by target-related information and target detection is accomplished through target identification. This seemingly natural view was in fact entailed in Neisser's model of search control (Neisser Reference Neisser1967). Based on Selfridge's Pandemonium architecture (Selfridge Reference Selfridge1959), the model claimed that task implementation activates stimulus analyzers for targets, but not for distractors, yielding deep target processing but only shallow distractor processing. As a result, distractors remain unidentified whereas target detection is brought about by target identification.
However, there is now substantial evidence suggesting that this view is fundamentally mistaken. The evidence applies to both target detection and distractor examination (for overviews, see Prinz Reference Prinz and Dornič1977; Reference Prinz1986).
First, target detection does not require deep modes of target processing. Targets are rather detected without being identified. This conclusion is suggested by several converging observations. (1) Pure detection: In tasks requiring parallel search for several targets simultaneously, participants may detect upcoming target locations without/before knowing which target is actually present. Targets can thus be detected and localized before they can be identified. (Neisser Reference Neisser1963; 1967, p. 100). (2) Pseudotarget detection: When a novel distractor appears in the search array, it is often treated as a target. Novel distractors are thus detected without being searched for (Prinz Reference Prinz1979, Exp. 1; Prinz et al. Reference Prinz, Tweer and Feige1974). (3) Detection at distance: Targets are often detected prior to being fixated. They can thus be detected without being the focus of spatial attention (Prinz Reference Prinz, Groner, Menz, Fisher and Monty1983; Prinz & Kehrer Reference Prinz, Kehrer, Groner and Fraisse1982). In sum, these observations suggest that targets are detected by default, not by identification. Target detection seems thus to rely on representational resources other than selectively primed target templates.
Second, distractor examination does not rely on shallow modes of distractor processing. Distractors are rather processed at considerable depth. This conclusion is suggested by a further set of observations: (1) Distractor complexity: Array complexity (as defined by the size of the distractor set) is one of the most powerful drivers of search performance (Prinz Reference Prinz1979). (2) Distractor redundancy: Sequential redundancy of distractor strings and arrays has been shown to drive performance as well. Importantly, distractor redundancy is automatically detected and exploited (Prinz Reference Prinz1979). (3) Distractor transfer: In paradigms requiring switches between search tasks, substantial transfer is obtained when the tasks share the same distractors (Prinz & Ataian Reference Prinz and Ataian1973). Remarkably, this also applies to visually different distractors that share the same names (Prinz et al. Reference Prinz, Hartlich and Lahmeyer1972). In sum, these observations suggest that distractor-related information is not discarded at all. Instead, it seems to be automatically identified within fixations, updated and integrated across fixations, and stored across consecutive trials.
Taken together, these findings suggest a framework for task implementation and top-down control that is actually the inverse of Neisser's intuitive model (Prinz Reference Prinz and Dornič1977; Reference Prinz1986; for overviews of more recent evidence on information integration and task-driven control see Schneider, Reference Schneider2013; Schneider et al. Reference Schneider, Einhäuser and Horstmann2013). This framework builds on three major claims: (1) Incoming information is continuously processed and integrated over space and time (i.e., within and across fixations, respectively). (2) Integration processes generate, maintain and update a dynamic forward model of current and impending distractor information (simultaneously for features, items, strings, etc.). (3) Once such a model is in place subsequent fixation samples are tested against it. As long as they match it, the scan is continued (= pure distractor samples). However, when a sample fails to match it, the scan gets disrupted for closer examination (= mixed distractor/non-distractor samples).
Regarding target processing, this framework accounts for detection by default. Targets are oddballs that fail to match the dynamic distractor model, be it at feature, item, or string level. Regarding distractor processing, it accounts for task implementation and top-down control. Task implementation arises as a by-product of automatic distractor processing: Continuous integration of distractor information within and across fixations generates a dynamic distractor model that instantiates the control structure for the current task. Top-down control is then based on using this model for testing upcoming fixation samples.
These are lessons from continuous search. Why should we take them to heart? On the one hand, continuous search differs from discrete search in several important respects so that only some of these lessons may be directly applicable to that domain. On the other hand, continuous search must be considered a powerful paradigm in its own right. It addresses processing mechanisms for active, extended search as they underlie natural search episodes in real-world scenes and settings.