Statistical learning of across-trial regularities during serial search.

Abstract

Previous studies have shown that attention becomes biased toward those locations that frequently contain a target and is biased away from locations that have a high probability to contain a distractor. A recent study showed that participants also learned regularities that exist across trials: Participants were faster to find the singleton when its location was predicted by the location of the target singleton on the previous trial. Note, however, that this across-trial statistical learning was only demonstrated for parallel search involving “pop-out” singleton targets. The current study investigated whether there is also learning of across-trial regularities when search is serial, using a T-among-Ls task. In Experiment 1, using search displays with a gray T-target among gray Ls, we found that participants did not learn the existing across-trial regularities. In Experiment 2 we used the same display and same regularities except that during the first half of the experiment the targets were colored red, allowing feature search. Critically, now participants did learn the across-trial regularities during pop-out feature search and the learned biases persisted when search was serial again. Participants were not aware of these regularities suggesting that learning was automatic and implicit. We propose that across-trial target-target associations learned during feature search shape a flexible priority map whereby the selection of the predicting location results in up-weighting of the predicted location on the next trial. This flexible priority map remained active even when search task changed dramatically from parallel to serial search.

What to expect where and when: how statistical learning drives visual selection

“…For example, in 'probability cuing' studies [29][30][31]58] in which a target appears more often in one region than in other regions, each time attention is directed to the target, associations are formed between the target and its location within the visual field. In a recent study investigating across trials learning it was shown that if there is no initial direction of attention to the target (for example, when search is serial rather than parallel), learning across trial regularities does not occur, but that learning can be instantiated when targets are made salient such that they pop out from the display [59]. We assume that spatial statistical learning operates by continuously adjusting weights within an assumed spatial priority map, which at any moment in time dynamically controls the deployment of covert attention and gaze [20].…”

“…Even though these regularities were randomly mixed within random trial sequences, the visual system was nevertheless sensitive to these regularities and adjusted the priority landscape on a trial-by-trial basis. As indicated, a follow-up study showed that in order to learn these across-trial regularities, the target must be salient and needs to pop out from the display [59].…”

“…Learning where and when Humans do not only learn to expect particular locations across trials [59,77], they also learn to orient attention to particular locations at specific moments in time [80,81]. In a pioneering modification of the Posner's orienting task by Coull and Nobre [82], instead of cueing the likely target location, cues predicted the time interval after which a target would likely occur.…”

No evidence for spatial suppression due to across-trial distractor learning in visual search

“…In more recent years, VSL was also investigated in the context of visual search (see Theeuwes et al, 2022 , for a recent review). It was shown that temporal predictive associations regarding spatial configurations and target locations in search displays can be implicitly learned and utilized to facilitate search (Boettcher et al, 2022 ; Li et al, 2022 ; Li & Theeuwes, 2020 ; Olson & Chun, 2001 ; Ono et al, 2005 ; Thomas et al, 2018 ; Toh et al, 2021 ). For instance, in the study by Li and Theeuwes ( 2020 ), participants were asked to search for a shape singleton target within a circular array of eight items (i.e., a diamond among seven circles or a circle among seven diamonds).…”

“…It was reasoned that if a particular location was selected (the predicting location), the weight of the predicted (i.e., expected) location on the next trial was boosted, resulting in a faster selection of the predicted location. In a follow-up study using the T -among- L s task as an operationalization of slow inefficient serial search, Li et al ( 2022 ) found that individuals did not express any RTs benefits on predicted trials. However, when observers were first exposed to the same regularities during parallel “pop-out” feature search (i.e., the target was colored differently from nontargets), across-trial VSL effects reoccurred, and the learned biases even persisted when search subsequently became inefficient and serial.…”