The long and the short of priming in visual search

Memory affects visual search, as is particularly evident from findings that when target features are repeated from one trial to the next, selection is faster. Two views have emerged on the nature of the memory representations and mechanisms that cause these intertrial priming effects: independent feature weighting versus episodic retrieval of previous trials. Previous research has attempted to disentangle these views focusing on short term effects. Here, we illustrate that the episodic retrieval models make the unique prediction of long-term priming: biasing one target type will result in priming of this target type for a much longer time, well after the bias has disappeared. We demonstrate that such long-term priming is indeed found for the visual feature of color, but only in conjunction search and not in singleton search. Two follow-up experiments showed that it was the kind of search (conjunction versus singleton) and not the difficulty, that determined whether long-term priming occurred. Long term priming persisted unaltered for at least 200 trials, and could not be explained as the result of explicit strategy. We propose that episodic memory may affect search more consistently than previously thought, and that the mechanisms for intertrial priming may be qualitatively different for singleton and conjunction search. Electronic supplementary material The online version of this article (doi:10.3758/s13414-015-0860-2) contains supplementary material, which is available to authorized users.


SAM as a model of visual priming by episodic retrieval
We used SAM as described in (G.-J. Mensink & Raaijmakers, 1988;J. G. W. Raaijmakers, 2003) and applied it to a typical visual search priming study. However, there are differences between the tasks that SAM was designed for ('memory tasks') and the way episodic retrieval may underly visual priming, which we took into consideration for our simulations. For example, memory tasks tend to have separate study and test phases which have participants actively study the items and 'search' for them in memory, respectivelytwo processes which can be assumed to be absent -or at least mostly implicit and passiveduring a visual search task. Secondly, memory tasks involve a multitude of memoranda that can be easily individuated, and are recalled separately (e.g. in the paired associates task, the cue a − −? has only one correct associate b). In the priming tasks simulated here, there are only two trial types (here: 'red' and 'green' trials). Without a well-defined individuating cue, so the retrieval of both trial types may simultaneously be probed throughout the trial 1 . Retrieval of memory traces laid down by any one of these two trial types will similarly have only one of two effects: the priming of either red or green targets. Finally, memory tasks challenge the participant to retrieve many different memoranda after a long delay. In priming in visual search, it is unlikely that participants will not be able to remember only two frequently presented target types. Therefore, it seems that for priming, the relative memory strength or ease of retrieval of the target types is crucial.
To simulate priming in visual search, we used the SAM model as outlined in (G.-J. Mensink & Raaijmakers, 1988; J. G. W. Raaijmakers, 2003) while accounting for these considerations: • all parameters (Table 1) were given values adopted from previous SAM papers except one: since learning was no longer embedded in a study phase, we decreased the learning rate (w).
• we simulated the memory strengths of only two items, representing the two trial types ('red' or 'green') which were associated with the fluctuating context whenever they were presented. This allowed us to simulate priming for either trial type through the respective memory strengths.
• As a measure for priming effects caused by retrieval of memory traces, we computed the probability of retrieval for both memory images. Then we defined the amount of facilitation by a memory item (F i ), as its recall probability divided by the sum of both recall probabilities. Since there are no explicit retrieval cues, this probability of recall is determined by the contextual cues, and priming is thereby determined by the bias in memory. Note that this measure is very similar to 'global familiarity' as was introduced to simulate recognition tasks with SAM (Gillund & Shiffrin, 1984).
1 One could argue that the onset of a trial constitutes a retrieval cue, or that other aspects of a trial probe retrieval, for example the stimulus layout. The latter certainly seems to be the case with the contextual cueing effect (Chun & Jiang, 1998, ,discussed in the main article discussion). Similarly, the episodic retrieval account explains 'episodic priming' effects through better retrieval when all visual features match the previous trias compared to when they do not (Hillstrom, 2000;Huang, Holcombe, & Pashler, 2004). For the present simulations, however, we focus on priming of pop-out tasks (Maljkovic & Nakayama, 1994) where trials only differ in their color, and all other aspects of the tasks are balanced. • We had no explicit hypothesis how memory retrieval influences RTs throughout the search trial. We assumed that this influence could be task-and participant-specific, and that the influence would saturate at a certain level. Therefore we fitted the following function to experiment data: Where c reflects a baseline RT when priming is maximally effective, assuming that memory influence can not facilitate RTs below a asymptotic minimum; g scales the priming effect, and the exponential term describes how fast priming saturates (τ ). Note that Figure 1 in the article depicts F i rather than simulated RTs.
We simulated data from three experiments: • Maljkovic and Nakayama (1994, Figure 7, bottom): Maljkovic and Nakayama reported the average RTs of each trial on which the target had the same versus a different color N trials in the past. The data is from one naive participant, and illustrates how priming gradually decays over the course of multiple trials. The simulations reveal the same pattern, although priming decays somewhat faster.
• Brascamp, Pels, and Kristjánsson (2011, Figure 1D): the time course of priming was probed by exploring how facilitation 'accumulates' over multiple repetitions of samecolored build-up trials, then 'breaks down' over intervening trials of the other color. measured by the RT on one final trial of the build-up color. Data comes from from six participants. Especially with few intervening trials, different build-up conditions show different priming.
• Martini (2010, Figure 5, left): Martini formalized short-term decay in priming, combining z-scored RTs from 50 participants, and computing the average contribution of trials matching in color N trials back. The facilitation (through priming) of each trial in the past again decays over several trials. The SAM simulations very closely match the empirical data, illustrating that both models are comparable under these conditions where target types are balanced.
All simulated experiments were repeated 25 times to reflect multiple participants, and all runs were preceded by a 'training phase' where 20 trials -10 of each type -were presented in random order. The simulations of the data from Maljkovic and Nakayama; Martini were then followed by 500 trials, balanced for both colors, in random order. For the data of Brascamp et al., each combination of buildup and intervening trials was ran, separately. The plots of the experiment data and the SAM-simulation data are given in figure 1. All three experiments were simulated very well by the model.

Raw Response Times
The plots in the paper all reflect color corrected response times. Here, we visualize the average raw RTs, in Figure 2.  (Maljkovic & Nakayama, 1994). B the average RT of trials of one particular color after a specified number of 'buildup' trials of the same color and a number of 'intervening' trials of the opposing color (Brascamp et al., 2011). C the facilitative effect evoked by a trial repetition from n trials in the past, computed from z-scored response times from 50 participants (Martini, 2010).