Visual Search as a Combination of Automatic and Attentive Processes

Visual Search as a Combination of Automatic and Attentive Processes Chris Donkin (cdonkin@indiana.edu) Rich Shiffrin (shiffrin@indiana.edu) Department of Psychological & Brain Sciences, 1101 E. 10 th St. Bloomington, IN 47408 USA Abstract We present a model in which visual search behavior is assumed to result from a combination of controlled, serial search and automatic attraction of attention to target stimuli. The model provides a quantitative framework for how these different processes are combined, and despite a large number of constraints, it is highly successful in accounting for human search behavior at the level of full response time distributions and choice probabilities. Keywords: visual search; computational modeling; automatic and attention processes; response times. Visual search tasks usually require an observer to determine whether or not a pre-defined target object is present within a display of objects (the display size, D, is the number of objects in the display, and displays either contain all foils, or instead one target with the rest foils). Performance in such tasks is usually measured by response time, because accuracy tends to be quite high. The results have been used to understand the processes of visual search, the factors that determine attention allocation, and the use of automatic and parallel vs. controlled and serial processes of comparison. The last of these is the focus of this research. Various factors contribute to whether search is automatic or controlled. Low-level perceptual differences between targets and foils, for instance a green target amidst red foils, facilitate automatic search behavior, produce response times that do not vary (much) with display size, and are often termed to ‘popout’ in line with subjective impressions (Triesman & Gelade, 1980). Even when targets do not ‘popout’ perceptually, consistent training (in which targets remain targets, and foils remain foils) in most cases gradually causes the targets to attract attention automatically, measured by the fact that the dependence on display size drops (Shiffrin & Schneider, 1977). However, the amount of such learning is dependent on the relation of the target shapes to foil shapes: For example, when a conjunction of features is needed to define targets, search generally appears more controlled (slower search rates) and automatic attraction of attention to targets is slower to develop. When the plot of mean response time, RT, to D has a slope near zero, search is considered parallel and automatic; when it has a large slope (usually roughly linear) search is often assumed to be serial (one item in the display at a time), or controlled. If the slope of target absent responses is about twice that for target present responses search is usually assumed to terminate once a target is found in the serial set of comparisons. Townsend and colleagues have argued convincingly, however, that analyses based on mean response times in standard visual search are relatively uninformative regarding the processes underlying search (e.g. Townsend & Nozawa, 1995). For example the function relating mean RT to D is insufficient for distinguishing serial from parallel search without strong additional assumptions. As we shall see, much more can be learned about visual search through analysis of full response time distributions. Complementing research identifying conditions in which search might be either automatic or serial (Thornton & Gilden, 2007), is research developing a unified model which can account for all search behavior (e.g., Wolfe’s guided search model, 2007). Our aim, in what follows, is to build upon these efforts and develop and implement a framework for how automatic and controlled search processes combine to produce the various types of observed visual search behavior. The model is fit to results partially reported in Cousineau and Shiffrin (2004), in which three participants received up to 80 sessions of training. It is similar to Wolfe’s Guided Search theory in a number of respects, including a serial search process that is guided to target items by an automatic parallel process. Model When a display appears a set of consecutive serial comparisons without replacement is initiated. The order of comparisons is chosen by the observer, and is random with respect to the actual target position. The order of comparisons can be interrupted, however, when search is guided by a separate parallel process that forces the next comparison to a target position. Search terminates when a comparison to a target occurs with a positive (i.e. target present) response, or with a negative response (i.e. target absent) when all display positions are compared unsuccessfully. As we will discuss later, however, a separate decision is sometimes made to terminate before all comparisons are finished (i.e. early termination of search). The Serial Comparison Process Each comparison involves a decision as to whether an item in the display is either a target or a foil. We model this decision using a relatively simple evidence accumulation model, based on the Linear Ballistic Accumulator (LBA) model (Brown & Heathcote, 2008). Figure 1 contains a graphical depiction of the comparison process. We assume