Learned predictiveness acquired through experience prevails over the influence of conflicting verbal instructions in rapid selective attention

Previous studies have provided evidence that selective attention tends to prioritize the processing of stimuli that are good predictors of upcoming events over nonpredictive stimuli. Moreover, studies using eye-tracking to measure attention demonstrate that this attentional bias towards predictive stimuli is at least partially under voluntary control and can be flexibly adapted via instruction. Our experiment took a similar approach to these prior studies, manipulating participants’ experience of the predictiveness of different stimuli over the course of trial-by-trial training; we then provided explicit verbal instructions regarding stimulus predictiveness that were designed to be either consistent or inconsistent with the previously established learned predictiveness. Critically, we measured the effects of training and instruction on attention to stimuli using a dot probe task, which allowed us to assess rapid shifts of attention (unlike the eye-gaze measures used in previous studies). Results revealed a rapid attentional bias towards stimuli experienced as predictive (versus those experienced as nonpredictive), that was completely unaffected by verbal instructions. This was not due to participants’ failure to recall or use instructions appropriately, as revealed by analyses of their learning about stimuli, and their memory for instructions. Overall, these findings suggest that rapid attentional biases such as those measured by the dot probe task are more strongly influenced by our prior experience during training than by our current explicit knowledge acquired via instruction.


Introduction
Attention and predictive learning are intimately related in a bidirectional way. On the one hand, we learn more from attended stimuli than from unattended stimuli that are present concurrently in the environment [1][2][3]: That is, attention influences learning. On the other hand, learning about the predictiveness of stimuli has been shown to play an important role in PLOS [7,8,11]. Hence it remains possible that initial, rapid attentional orienting is influenced primarily by experience of predictiveness (i.e., by selection history), but that this initial experience-driven bias is subsequently overridden by the influence of explicit knowledge-and it is this latter factor that dominates in Mitchell et al.'s dwell time measure. Notably, Mitchell et al.'s dwell time measure summed gaze over a relatively long period (1.25 sec) and hence any rapid influence of selection history may be dominated by a slower effect of explicit knowledge when integrated over the whole period. Moreover, whereas attention and eye movements are generally quite tightly coupled [12], it is possible for rapid shifts of attention to occur covertly; that is, in the absence of eye movements [13]. Such covert attentional shifts would not be captured by eye-tracking. Thus it is possible that, even in the Change condition of Mitchell et al.'s study, there may have been a rapid (and possibly covert) attentional bias towards the stimulus that had been predictive during Phase 1, driven by selection history. This rapid bias may then have been followed by a second stage of overt attention to the previously-nonpredictive stimulus in line with verbal instructions that this stimulus would now be predictive, and it is this latter process that would be most evident in the dwell time measure used in this study.
The aim of the current study was to investigate the influence of experience of predictiveness (selection history) versus instructed knowledge on rapid-and potentially covert-shifts of attention. For this purpose, we used a dot probe task [14] as our measure of attention, based on previous research showing that this task provides a sensitive measure of rapid, predictiveness-related attentional bias. Le Pelley, Vadillo, and Luque in 2013 [14] (see also [15]) trained participants on a task in which a pair of stimuli (coloured shapes)-known as a stimulus compound-appeared on each trial, with one stimulus on the left side of the screen, and the other on the right. Participants had to learn to make one of two button-press responses. One of the stimuli presented on each trial predicted the correct response, while the other was nonpredictive, much as in the study by Mitchell et al. [10]. However, in this case attention to the stimuli was measured using a dot probe task, which is based on the idea that detection of a target will be faster if that target appears in an attended location than in an unattended location.
On each trial of the dot probe task in Le Pelley et al.'s study [14], participants were shown (briefly) one of the stimulus compounds that had been experienced during training. After a short stimulus-onset asynchrony (SOA) of 250 ms, a dot (the probe) could appear at the location of one of the two stimuli. Participants were required to respond to the appearance of the probe as quickly as possible. Importantly, across trials of the test phase the probe was equally likely to appear in the location of (that is, be cued by) the stimulus that had been predictive during the training phase as it was to be cued by the nonpredictive stimulus. Hence there was no advantage to be gained in directing attention to either location prior to probe presentation. Indeed, participants were explicitly informed that in order to respond to the probe as quickly as possible, their best strategy was to ignore the initially presented stimuli.
Despite this instruction, responses to the probe were significantly faster when it was cued by the predictive stimulus than when it was cued by the nonpredictive stimulus, suggesting that participants had rapidly oriented their attention to the location of the predictive stimulus prior to the appearance of the probe. Notably, Le Pelley et al. [14] demonstrated that providing more time for participants to process the stimuli-by increasing the SOA on dot probe trials to 1000 ms-significantly weakened the influence of predictiveness on dot probe responding. Consistent with the argument that we advanced earlier, these findings demonstrate that rapid attentional biases that can be detected at short SOAs might go undetected in tasks that measure the deployment of attention over longer periods of time, including on the timescale of the measure used by Mitchell et al. (~1 sec).
In general terms, our hypothesis is that rapid attentional bias towards previously predictive stimuli might be primarily determined by selection history, and relatively immune to the effect of instructions. To test this hypothesis, we conducted an experiment similar to Mitchell et al.'s Experiment 2 [10] but using a dot probe task to measure rapid-and potentially covert-attention to stimuli. During Phase 1, some stimuli were trained as predictive of the correct categorization responses while others were nonpredictive. During Phase 2, participants learned new categorization responses. Immediately before this second phase, participants received continuity or change instructions regarding which stimuli would be important in determining the correct response in the following phase. A dot probe task was combined with the learning task throughout the experiment, as in Le Pelley et al.'s Experiment 3 [14] (see also [5,16,17]). By analyzing response times to the dot probe during Phase 2, we could examine the impact of experienced predictivess provided through training (in Phase 1) versus instructions on attentional bias. Crucially, in the change condition, we predicted an attentional bias driven by experienced predictivess within the short SOA condition. In other words, despite the conflict between experienced predictiveness and instructions regarding which stimulus should be prioritised, the former factor would have a greater influence on attentional bias than the latter.

Materials and methods
The design of our study was conceptually similar to that of Mitchell et al. [10] in that it compared the influence of training versus instruction on predictiveness-related attentional biases. Our study departed from the procedure of Mitchell et al. by using a within-subjects manipulation of verbal instructions, in order to increase the sensitivity of the experiment (a similar approach was used in Don & Livesey's, Experiment 3 [18], and in Shone et al.'s Experiment 2 [19]). Accordingly, after Phase 1, participants were informed that four specific stimuli would be the most relevant to learn about during Phase 2. Participants then experienced different pairs of stimuli in Phase 2. In the consistent pair, instructions regarding relevance were consistent with the predictive or nonpredictive status of stimuli that had been experienced during Phase 1 training (see Table 1). In contrast, in the inconsistent pair, instructions regarding relevance were inconsistent with the status of stimuli established during Phase 1. Finally, we also included two pairs of novel stimuli in Phase 2 that had not appeared in Phase 1. One stimulus of each pair was instructed as being relevant in Phase 2, whereas the other was not. Since these stimuli had not undergone prior training, any attentional bias revealed in dot probe responding for these novel pairs can only reflect the influence of instructions (cf. [20]). Observing an attentional bias for novel pairs would also provide a manipulation check, showing that participants had read, understood, and followed the instructions regarding relevance prior to Phase 2.

Participants and apparatus
A total of 135 students from a Spanish university participated for course credit; 68 were randomly assigned to a short SOA group, and the remainder to a long SOA group. Written consent was obtained and the Human Research Ethics Committee of the University of Málaga approved the study. The experiment was carried out in a quiet room with 10 semienclosed cubicles each equipped with a standard PC and 38.4 cm monitor. The task was run using the Cogent 2000 toolbox (http://www.vislab.ucl.ac.uk/Cogent/) for MATLAB. Participants made all responses with the computer keyboard.

Stimuli
Stimuli were the same as those used by Luque et al. [16], and included eight equal-sized circles (diameter subtending 4.7˚visual angle at a viewing distance of~80 cm), with radiating lines of varying thickness (see Fig 1). These figures were filled with different, easily discriminable colours that had similar brightness. The (red, green, blue) values for each colour were: light redbrown (190, 86, 78), gold (190,185,78), green (93, 191, 77), turquoise (77,191,191), purple (132, 71, 255), pink (255, 5, 255), red (208, 0, 0), and grey (150, 150, 150). These stimuli were randomly assigned the roles indicated by letters A-H in Table 1. Additionally, there were four more white outline figures consisting of two identical rectangles, one horizontally and the other vertically oriented, and two identical ellipses, one horizontally and the other vertically oriented. These last figures were used for filler trials, and were assigned roles corresponding to letters I-L in Table 1. These stimuli were presented centrally in white square frames with sides subtending 6.4˚, which were located on the right and left sides of a small fixation cross that was located in the centre of the screen; the centre-to-centre distance between the two boxes subtended 6.4˚. The dot probe was a white square with side length subtending 1.1˚. This would appear superimposed centrally on one of the stimuli. The screen background was black.

Procedure
The procedure was similar to that described in Le Pelley et al.'s Experiments 2 and 3 [14] (see also [5,16]). Initial instructions (in Spanish) described the categorization task. Participants were told that, on each trial, a pair of stimuli would appear and they should make a categorization response by pressing either the '1' or '2' key with their left hand. Response keys '1' and '2' were randomly assigned the roles of response categories 1 and 2 shown in Table 1 for each participant. They were told they should try to learn the correct response for each pair of stimuli. Participants then underwent a first phase (Phase 1A) of 32 categorization trials. This comprised four eight-trial blocks, with each of the four stimulus pairs shown in Table 1 appearing twice per block in random order; for each stimulus pair, the predictive stimulus appeared once on the left and once on the right. On each trial a fixation cross appeared, followed after 500 ms by the pair of stimuli. After 1 s, a message framed within a central rectangle prompted participants to choose between response keys '1' and '2'. Incorrect responses were followed by the feedback message "Error! The correct response was [1/2]," which remained onscreen for 3 s; no explicit feedback was provided for correct responses.
Following Phase 1A, participants received further instructions explaining that on subsequent trials they would complete two tasks: On each trial (a) a pair of stimuli would appear; (b) a small white square (the dot probe) would then appear superimposed on one of these stimuli; (c) participants should press the left or the right arrow with their right hand depending on whether the square appeared on the left or on the right stimulus, respectively; (d) once they had responded to the square, they should make a categorization response to the stimulus pair using the '1' or '2' keys with their left hand as in the pretraining stage. Participants were told that they should respond to the position of the dot probe as rapidly as possible and that "In order to do so, it is best that you ignore the pair of figures until you have responded to the location of the square" (translated from Spanish). Fig 1 shows the event timing of a standard trial. Each such trial began with presentation of a central fixation cross. After 500 ms the stimulus pair appeared to either side of this cross. After an SOA of either 250 ms or 1,000 ms (depending on the SOA group to which the participant had been allocated), the dot probe appeared superimposed on one of the stimuli. This probe remained until participants made the correct response (left arrow key for a target presented on the left; right arrow key for a target on the right). Immediately on making the correct dot probe response, the probe disappeared and 1 s later the message "1 or 2?" appeared as for Phase 1A. Participants then made a categorization response using the '1' or the '2' keys; feedback was administered as in Phase 1A, and the next trial began after 1 s.
Participants completed Phase 1B, which comprised 10 blocks of 16 trials each (see Table 1). Each trial type of Phase 1B appeared four times; once for each combination of cue location (predictive cue on the left or on the right) and dot probe location (on the left or on the right stimulus). Therefore, the dot probe was equally likely to appear on the predictive or on the nonpredictive stimulus. The order of trials within each block was randomized.
Following Phase 1B, participants were told that in the next phase (Phase 2) they would learn new relationships between certain stimulus pairs and response categories 3 and 4 in a similar way as in Phase 1B. Some stimulus pairs had been presented in Phase 1A and 1B (which we refer to collectively as Phase 1), whereas others included new stimuli (see Table 1). Importantly, although all stimuli were in fact equally predictive of the response categories with which they were paired in Phase 2, participants were told that, from that moment on, the only relevant stimuli that they should use to choose the correct response key were A, D, E, and G. As explained in the Introduction, stimuli in Phase 2 were paired so as to create a consistent pair (AC) in which the instructed-relevant cue (A) had been experienced as being predictive in Phase 1; an inconsistent pair (BD), in which the instructed-relevant cue (D) had been experienced as being non-predictive in Phase 1; and two novel pairs (EF and GH), in which neither cue had appeared in Phase 1. Filler trials consisting of pairs IJ and KL were also included to increase the complexity of the learning task. The assumption underlying this procedural measure is that complex environments encourage the use of selective attention in order to focus and simplify information-processing. By increasing memory load in our critical test phase (Phase 2), we therefore hoped that these additional filler trials would provide additional drive for participants to deploy selective processes, e.g., by focusing on the cues mentioned in the verbal instructions. Phase 2 comprised four blocks of 24 trials each. Each of six stimulus pairs appeared four times per block, counterbalancing cue and probe location as in Phase 1B. Response categories 3 and 4 were randomly assigned to response keys '3' and '4' for each participant and independently of the assignment of response categories 1 and 2 to response keys '1' and '2'. Thus, these assignments were uncorrelated across participants.
After Phase 2, participants completed a judgment test phase in which they rated the extent to which each stimulus was associated with response categories 3 and 4, on a scale from 1 ('completely sure that Stimulus X does not predict Response Y') to 7 ('completely sure that Stimulus X predicts Response Y'). Participants rated each stimulus with regard to each of the response categories (3 and 4) in random order.
Finally, participants completed an instruction memory test to assess their memory for the instructions regarding which stimuli were relevant and which were not. Again, a rating scale from 1 to 7 was used, with 1 meaning 'completely sure that Stimulus X was not instructed as relevant', and 7 meaning 'completely sure that Stimulus X was instructed as relevant'. Participants provided ratings for all stimuli in random order.

Results
We imposed a selection criterion so as to exclude participants who did not show strong evidence of having learned the correct categorization responses by the end of Phase 1. Specifically, we excluded data from participants who failed to reach a criterion of 25 or more correct categorization responses in the two last blocks (32 trials) of Phase 1B since this constitutes strong evidence of learning (p = .001, binomial test). This resulted in exclusion of five participants from the short SOA group (final n = 63), and eight from the long SOA group (final n = 59).

Phase 1
Fig 2A shows the mean percentage of correct responses as a function of block and SOA group in Phase 1A (blocks 1-4) and 1B (blocks [5][6][7][8][9][10][11][12][13][14]. Participants' response accuracy increased over blocks; there was no apparent difference between SOA groups, with both approaching perfect accuracy during the final four blocks. These impressions were confirmed by a 14 (block) × 2 (SOA group: 250ms vs 1000ms) ANOVA, which yielded a significant main effect of block, F(13, 1560) = 99.5, p < .001, Z 2 p ¼ :45. Neither the main effect of SOA nor the block × SOA interaction was significant (Fs < 0.87). A follow-up analysis was conducted focusing on the last four blocks to test if both SOA groups were performing at a similar level by the end of Phase 1. This analysis found a marginally significant effect of block, F(3, 360) = 2.16, p = .093, Z 2 p ¼ :02; the main effect of group and the interaction between block and SOA were not significant (Fs < 0.58).
Following the same procedure as in Le Pelley et al. [14], response times (RTs) from the dot probe task were filtered and transformed before the analyses. First, we excluded RTs from trials in which the first response to the probe was an incorrect response (12.6% and 12.1% of trials in the short and long SOA groups respectively). Second, we excluded trials with RTs shorter than 150 ms and longer than 1500 ms: Responses outside these limits were a priori deemed to be anticipations or to reflect lack of focus of the task respectively. Next, log-transformation was applied to reduce the positive skew that is typical of RT distributions, such that the transformed data better fit a normal distribution. Finally, we screened the transformed data for outlying RTs on an individual basis, removing transformed RTs lying more than 3 SDs from each participant's mean. There is relatively large inter-subject variability in mean RT, such that for some (generally fast) participants there might be exceptionally slow responses (indicating inattention) that nevertheless have an RT below the 'hard limit' of 1500 ms. We applied the 3-SD criterion after log-transformation as the more normal distribution of transformed data means that SDs are more easily interpreted. Screening based on RTs resulted in removal of a further 4.7% and 4.6% of trials in the short and long SOA groups respectively.  [14], and indicate that predictive learning tended to produce an attentional bias towards the predictive stimulus. The fact that this bias was found in the 250 ms SOA condition but not in the 1000 ms SOA condition implicates a very rapid and short-lived attentional bias towards predictive stimuli.
As expected, the short SOA group showed an attentional bias towards stimuli previously learned to be predictive through trial-by-trial training. Crucially, this effect was not significantly affected by whether these stimuli had been explicitly instructed as relevant or not during Phase 2. In order to quantify the evidence in favour of a null effect of instruction on the attentional bias, we conducted a Bayesian analysis of this effect using JASP (version 0.9.0.1) [21], with the default Cauchy prior. The resulting Bayes Factor in favour of the null hypothesis over a one-sided alternative for the instruction effect within the short SOA group was BF 0+ = 8.74, i.e., substantial evidence for the null hypothesis [22]. Turning to the long SOA group, our ANOVA analysis suggested similar RTs regardless of the experienced predictiveness of stimuli or instructions. The fact that a significant attentional bias towards predictive stimuli was detected at short SOA but not long SOA is again consistent with the engagement of a fast attentional process that may go undetected if the attentional task does not impose strong enough time constraints.
One possible explanation of the failure of instructions to exert any significant effect on the data in Fig 3A is simply that participants did not read, understand, or make use of these instructions during Phase 2. To test this possibility, we analyzed the effects of instructions on RTs for novel stimulus pairs EF and GH. Recall that stimuli E and G were instructed as relevant during Phase 2, while F and H were noninstructed; none of these cues was experienced during Phase 1. Fig 3B shows mean log-transformed RTs during Phase 2. For these novel pairs, participants in the 250ms SOA group responded faster when the probe appeared on instructed stimuli than when it appeared on noninstructed stimuli. Participants in the 1000ms SOA group did not show a clear bias. A 2 (instructions: instructed vs noninstructed) × 2 (SOA), ANOVA yielded no significant effect (all Fs < 2.76, smallest p = .1). However, since our previous analyses suggest that attentional biases in the dot probe task were confined to the short SOA group, we used a t-test to analyze the effect of instructions on RTs within the short SOA group only. This revealed a significant effect of instructions, t(62) = 2.33, p = .023, η 2 = .08. This confirms that participants were effectively following the instructions about stimulus relevance, and that such instructions can produce a rapid attentional bias towards stimuli, at least when such instructions do not conflict with stimulus predictiveness experienced through trial-by-trial training.

Ratings of stimulus-outcome relationships
Participants' ratings from the judgment test were analyzed to assess the influence of experienced predictiveness and instructions on learning of stimulus-outcome relationships in Phase 2. Following Le Pelley and McLaren [3] (see also [23]), we calculated a rating score for each stimulus by subtracting the rating given to the incorrect response category from the rating Experienced predictiveness and attentional bias given to the correct response category. High, positive values on this scale (maximum = 7) indicate strong learning of a correct stimulus-outcome relationship. Fig 4 shows mean ratings for each stimulus (ratings for cues E and G, which were equivalent, were combined [denoted E/G]; ditto for cues F and H). The data relating to cues A-D were analyzed with a 2 (experienced predictiveness: Predictive vs nonpredictive during Phase 1) × 2 (instruction) × 2 (SOA) ANOVA. This revealed significant main effects of experienced predictiveness, F(1, 120) = 28.3, p < .001, Z 2 p ¼ :19, and instruction, F(1, 120) = 6.03, p = .015, Z 2 p ¼ :05. No other effects were significant (Fs < 1.51, smallest p = .221). Both short and long SOA groups learned more during Phase 2 about stimuli that had previously been experienced as predictive than those that had been experienced as nonpredictive. Both groups also learned more about stimuli that had been explicitly instructed as relevant during Phase 2 than those that had not been instructed. This latter finding once again confirms that participants read and made use of the instructions regarding relevance given prior to Phase 2.
Putting together the dot probe results from Phase 2 and participants' ratings of stimulusoutcome relationships for old stimuli, it seems that past experience with stimuli had an influence on rapid and short-lived attentional bias towards predictive stimuli, and on how much was learned about such stimuli in a subsequent phase of learning. By contrast, instructions about stimulus relevance had an effect on learning, as measured by subjective ratings, but not on rapid and short-lived attentional capture.
Regarding Stimuli E-H, there was a numerical trend towards higher ratings for the instructed stimuli than the noninstructed stimuli, but it did not reach statistical significance. A 2 (instruction) × 2 (SOA group) ANOVA on participants' ratings revealed no significant effects (all Fs < 1.55, smallest p = .217). Thus, in this case, instructions exerted an effect on attentional capture that did not translate into an advantage in terms of stimulus-outcome learning (as measured by explicit judgments). Fig 5 shows data from the final instruction memory test. The figure shows participants' ratings of their confidence in their memory regarding the instructions that accompanied each stimulus prior to Phase 2 (1 = sure that the stimulus was not instructed as relevant; 7 = sure that the stimulus was instructed as relevant). Instruction memory ratings are shown as a function of whether stimuli were actually instructed as relevant or not, SOA group, and compound consistency. This last factor refers to the type of stimulus compound in which stimuli appeared during Phase 2. Table 1 shows that for stimuli A and C (which appeared together in the AC compound) experience of predictiveness during Phase 1 was consistent with instructed relevance prior to Phase 2. In contrast, for stimuli B and D (appearing in the BD compound), instructed relevance was inconsistent with participants' prior experience of the predictiveness of these stimuli (e.g., B was experienced as predictive during Phase 1, but was instructed as irrelevant in Phase 2). Finally, stimuli E-H were new in Phase 2: the predictiveness of these stimuli had not been established prior to instructions. Hence the compound consistency factor highlights the extent to which instruction memory was affected by the congruency between participants' experienced predictiveness, and the instructions they received. A 2 (instructions) × 3 (compound consistency: AC vs BD vs EF/GH) × 2 (SOA) ANOVA on participants' instruction memory ratings revealed significant main effects of instruction, F(1, 120) = 64.85, Overall, instruction memory ratings were higher for cues that were instructed as relevant than for those that were not instructed, confirming again that participants had read and remembered these instructions. Interestingly, however, the effect of instructions differed as a function of stimulus type: Fig 5 suggests a larger

Discussion
The primary aim of the current experiment was to assess whether Mitchell et al.'s [10] results -a complete reversal of the effect of experienced predictiveness on selective attention due to instructions about stimulus relevance-could be replicated if predictiveness-driven attentional bias was assessed through a dot probe task instead of an eye-tracking technique. To this end, participants experienced differences in the predictiveness of different stimuli over the course of trial-by-trial training in a first learning phase, and, later on, received verbal instructions regarding stimulus relevance for the subsequent learning phase that could be either consistent (AC compound) or inconsistent (BD compound) with experienced stimulus predictiveness. We measured the effects of these manipulations on spatial cueing in the dot probe task following the same procedure as Le Pelley et al.' Experiment 3 [14]. Like Le Pelley et al. [14] (see also [15]), the current experiment found that-with a short stimulus-onset asynchrony (SOA) between the stimuli and the probe-responses to the probe during Phase 2 were faster when its position was cued by stimuli previously experienced as predictive compared with nonpredictive stimuli. This suggests that experienced predictiveness produced an attentional bias towards predictive stimuli. The fact that experienced predictiveness produced a bias in spatial cueing of the probe only at short SOA (250 ms) and not at longer SOA (1000 ms), suggests the operation of a rapid and short-lived attentional process.
Most importantly, the rapid attentional bias towards predictive stimuli (observed at short SOA) was not reversed or even significantly altered by conflicting verbal instructions regarding stimulus relevance, with a Bayesian analysis suggesting substantial support for the null hypothesis. This was not due to participants' failure to understand, retrieve, and follow verbal instructions. First, instructions regarding stimulus relevance affected explicit ratings about stimulus-outcome relationships learned in Phase 2. These ratings clearly show that participants tended to learn more about stimuli instructed as relevant (A & D) than noninstructed stimuli (B & C). Second, instructions produced an attentional bias towards new stimuli instructed as relevant (E & G) relative to new stimuli that were noninstructed (F & H). Finally, memory for instructions was reasonably good as evidenced by participants' higher instruction-memory ratings to stimuli instructed as relevant than noninstructed stimuli.
Thus, despite evidence that participants had read, understood, and implemented verbal instructions regarding stimulus relevance, these instructions had no effect on the bias in rapid attentional orienting to stimuli that had previously been experienced as predictive, compared to those experienced as nonpredictive. This suggests that trial-by-trial experienced predictiveness (i.e., selection history) drives the development of a rapid and relatively inflexible attentional bias that is somewhat insulated from changes in explicit knowledge about predictive status produced by verbal instructions. Note that we are not claiming here that performance in the dot probe task at short SOA is generally immune to verbal instructions. Indeed, our own data suggest this is not the case-for the novel stimuli (that had not been experienced during Phase 1), responses to the dot probe were significantly faster when it was cued by a stimulus that had been instructed as relevant (E/G) than when it was cued by a stimulus that had not been instructed (F/H) (for related findings, see [24,25]). The novel finding of our data is that the influence of prior experience of predictiveness on rapid attentional bias is sufficiently strong that, given a difference in selection history, no effect of attentional control via instruction is observed.
In line with previous evidence [10,18,19], we found that participants' learning of stimulus-outcome relationships during Phase 2 was influenced by instructions regarding relevance: Participants learned more, in general, about stimuli instructed as relevant than those that were not instructed. That said, the influence of instructions on learning was relatively slight, and was not sufficient to overcome the influence of experienced predictiveness on learning. That is, we also observed a main effect of experienced predictiveness on participants' judgments of stimulus-outcome relationships, and instructions were not sufficient to reverse the pattern of greater learning about stimuli experienced as predictive than those experienced as nonpredictive. This is indicated by the finding that stimulus D (experienced as nonpredictive but instructed as relevant) produced weaker judgments than stimulus B (experienced as predictive but not instructed as relevant). Thus while demonstrating an influence of verbal instructions about stimulus relevance on learning, our data fail to replicate Mitchell et al.'s finding of a complete reversal of the effect of experience as a result of instructions [10]. In this respect our data are more similar to subsequent findings that have also failed to replicate this full reversal [18,19]. Taken together, these findings suggest that both selection history produced via repeated experience with stimuli, and verbalisable knowledge, may contribute to biases in learning towards predictive cues observed in earlier studies (e.g., [2,3,26]).
We implemented instructions regarding stimulus relevance by explicitly informing participants which specific stimuli would be relevant during Phase 2 (following a procedure used by Don & Livesey in 2015 [18], and by Shone et al. in 2015 [19). This differed from the approach used by Mitchell et al. [10], who provided the more general instruction that stimuli which had been predictive during Phase 1 were highly likely (in the Continuity condition) or highly unlikely (in the Change condition) to be predictive during Phase 2. It seems unlikely that this procedural difference was responsible for the persistent, rapid attentional bias towards stimuli experienced as predictive observed in the dot probe task of the current experiment. As Don and Livesey [18] noted, the instructions used by Mitchell et al. [10] might actually result in a rapid attentional bias towards stimuli previously experienced as predictive even in the Change condition, since participants may first need to identify the stimulus that was previously predictive in order to identify the stimulus which was previously nonpredictive (and which should now be attended, according to instructions). In contrast, direct instruction regarding which cues are relevant in Phase 2 does not require that participants first identify the stimulus which used to be predictive in Phase 1. Consistent with this claim, Don and Livesey [18] showed that instructing the relevance of specific stimuli results in, if anything, a larger influence of instructions on stimulus-outcome learning than does providing more general instructions regarding continuity/change, as used by Mitchell et al. [10]. This implies that the procedure used in the current experiment should have been at least as sensitive to showing an effect of instructions on attentional orienting as that used by Mitchell et al., if such an effect were to exist.
Although it is not possible to draw strong conclusions about the nature of the attentional processes underlying the effects found in our experiment, we briefly consider this issue in this final section. As noted in the Introduction, given the specific characteristics of the dot probe task used here, especially in the short SOA condition, it is reasonable to think that this task may be more sensitive to rapid and relatively inflexible mechanisms of selective attention that tend to persist despite updates in current explicit knowledge and goals, as compared to the eye-tracking measure used by Mitchell et al. [10]. Accordingly, one interpretation of our results is that the rapid and short-lived influence of selection history reflects a relatively automatic process over which participants have little strategic control (cf. [8,9,27]). On this account, repeated experience of attentional selection of a particular stimulus produces an automatic and habitual prioritization of that stimulus. In the current dot probe task, the locations of the predictive/nonpredictive stimuli were noninformative with regard to the location in which the probe would appear. Considering this task on its own, then, there was no advantage to be gained in strategically directing attention to either location prior to the onset of the probe-the implication being that the observed attentional bias towards predictive stimuli did not reflect strategic allocation of attention, but rather an involuntary process. The long SOA condition may then have provided sufficient time for a more strategic, top-down attentional process to return attention to the centre of the display.
However, an alternative account is possible. Notably, the dot probe task was embedded within predictive learning trials in this experiment, and this overlap in task structures raises questions over the strategies that participants might have used. In particular, while participants were instructed to ignore the stimuli until after they had responded to the dot probe, they may nevertheless have begun a strategic process of identifying the stimuli and preparing a categorization response prior to the onset of the probe. On this account, then, the rapid attentional bias towards predictive stimuli demonstrated in the dot probe task may result from a voluntary process. The absence of a bias at long SOA might then be because 1000ms provided sufficient time for participants to program a categorization response and then return attention to the centre of the display in anticipation of the upcoming dot probe. Additionally the fact that RTs in the short SOA group were longer than in the long SOA group may also be seen as consistent with the idea that participants spent time preparing for a categorisation response before responding to the dot. According to this, the effect of SOA on RTs may be seen as a typical case of cognitive bottle neck in concurrent multitasking preparations (see [28], for a review on this issue). Note, however, that this effect of SOA on participants' RTs has also been found even when the learning and the dot probe tasks take place in separate trial blocks [14].
Thus we have two alternative accounts: One which invokes opposing involuntary and strategic attentional processes, and the other in which allocation of attention is entirely strategic. The current findings do not allow us to decide between these alternatives (though we note that influences of experienced predictiveness on dot probe performance can be observed even when the two tasks are entirely separate, which is harder to reconcile with the wholly-strategic account; see Experiment 2 in [14]). For current purposes this issue is not critical, however: The important finding is that the processes underlying the influence of learned predictiveness on attention show distinct influences of selection history and explicit knowledge. This is true whether we align selection history with involuntary and explicit knowledge with voluntary attention, or whether selection history and explicit knowledge both exert distinct effects on strategic orienting. Having established a distinction here, future studies could further investigate the nature of the underlying cognitive processes.