The sensitivity of both the detection and discrimination tasks to location probability learning raises an important question about the transferrability of learning. If changes in spatial attention following training readily transfer across tasks, then the design of training tasks may be guided by convenience. For example, any stimuli and tasks might be used for training in medical imaging, as long as the spatial regularities are maintained. However, studies reviewed earlier suggest that location probability learning does not always transfer between tasks, especially if one of the two tasks does not involve visual search (e.g. treasure hunt or scene memory). Although both the detection and discrimination tasks used in this study involved visual search, differences in how well search items could be segmented may influence how people shift attention in these tasks.
Participants in Experiment 2 were randomly assigned to acquire location probability learning in either the detection or the discrimination task. Unlike Experiment 1, the task changed in the testing phase: from detection to discrimination or vice versa. We examined whether location probability learning acquired in one task transferred to the other.
We included eye tracking for a subset of the participants. This measure yielded insights into potential differences in how search was conducted. We examined whether the two tasks involved serial search (e.g. participants make multiple fixations before finding the target) and, if so, whether they differed in the number and duration of fixations. Eye tracking also provided an additional measure of a search habit: the direction of the first saccadic eye movement. Previous studies showed that location probability learning not only facilitated RT, but also increased the proportion of first saccades toward the high-probability quadrant (Jiang, Won, & Swallow, 2014; Salovich et al., 2017). Discrepancies sometimes occurred, however. The first-saccadic preference emerged more slowly than the RT advantage (Salovich et al., 2017). These findings suggest that covert attention – attentional shifts without eye movements – rely on similar, but not identical, mechanisms as overt shifts of attention with eye movements. Differences between the two raise the possibility that RT and first saccades may show different patterns of learning and cross-task transfer, a possibility tested in Experiment 2.
Method
Participants
Sixty-four college students completed Experiment 2. All participants were drawn from the same participant pool. The first 32 participants were tested without an eye tracker. Among them, a random half were trained in the detection task and tested in the discrimination task, whereas the task assignment was reversed for the other half. Eye tracking was added for the last 32 participants. A random half of these participants were trained in the detection task and tested in the discrimination task and the other half were assigned the opposite tasks. Altogether, 32 participants completed the detection training (26 women and six men, mean age 20.0 years) and 32 participants completed the discrimination training (26 women and six men, mean age 20.5 years).
Procedure and design
Similar to Experiment 1, participants first underwent a thresholding task to determine the noise opacity for the detection task and the target-distractor similarity for the discrimination task. Thresholding was done on both tasks in separate blocks, counterbalanced in order between participants. The mean noise opacity level used in Experiment 2 was 93% in the detection task. The mean similarity level was an offset of 24 pixels.
Next, the detection training group carried out the detection task in seven blocks, then switched to the discrimination task for four blocks. The discrimination training group carried out the discrimination task in seven blocks, then switched to the detection task for four blocks. In both groups, the seven training blocks involved a biased target distribution: the target, when present, appeared in a high-probability quadrant on 50% of the trials and in each of the other quadrants 16.7% of the trials. The last four testing blocks involved an unbiased target distribution: the target, when present, appeared in each quadrant 25% of the time. This experiment was comparable in design to that of Experiment 1. The key difference is that the task changed between training and testing. Recognition test was conducted at the completion of the visual search task.
Eye tracking
Eye-tracking participants rested their head on a chinrest. An EyeLink 1000 eye tracker (SR research Ltd., Mississauga, ON, Canada) tracked the left eye at a sampling rate of 2000 Hz. Eye position was calibrated before the experiment and verified before each trial. Recalibration was done as needed. The eye tracker recorded the eye position and information about saccades and fixations. We focused on: (1) the number of fixations per trial; (2) the duration of each fixation; and (3) the direction of the first saccadic eye movement after trial onset.
Statistical analysis
In addition to repeated measures ANOVA, we performed a Bayesian analysis on the testing phase data. In the case of a null effect, the Bayesian analysis tests whether a lack of an effect is more plausible than the presence of an effect. This Bayesian analysis was implemented in the BayesFactor package in R (Rouder & Morey, 2012; Rouder, Morey, Speckman, & Province, 2012). We used the default prior (Cauchy prior) in this package, which has been shown to be appropriate for the vast majority of designs in experimental psychology (Rouder et al., 2012). We used a top-down model comparison to assess the evidence for or against probability cuing in the testing phase. This procedure first constructed a full model including all terms. Next, it took out each term one at a time and compared the resulting model with the full model. Each term yields a Bayes Factor, which describes the degree to which the model omitting that term is preferred over the full model. For example, a Bayes Factor of 5 implies that a model omitting the term is five times more plausible than a model including it. In other words, it is five times more likely that term does not have an effect than it does.
Results
Behavioral data
Behavioral data were obtained from the whole sample. Accuracy was 98.8% on target absent trials (false alarm rate 1.2%). On target-present trials, accuracy was unaffected by the target’s quadrant. It was 89.1% in the high-probability quadrant, 88.4% in the low-probability quadrants, t(15) = 1.06, p = 0.30 for the detection-training participants; 90.9% in the high-probability quadrant, 89.2% in the low-probability quadrants, t(15) = 1.78, p = 0.09 for the discrimination-training participants. RT was longer on target-absent than target-present trials (5698 ms vs 1758 ms for the detection-training participants and 4784 vs. 2471 ms for the discrimination-training participants). We examined mean RT from correct target-present trials, excluding outliers (< 250 ms: 0.01% of the trials; > 10 s: 0.27% of the trials). Figure 3 displays these results.
The training phase was the same as in Experiment 1. Replicating Experiment 1’s finding, we found significant location probability learning. An ANOVA using task, target quadrant, and training block as factors showed a significant main effect of target quadrant, as RT was faster when the target was in the high-probability quadrant, F(1, 62) = 47.90, p < 0.001, ηp2 = 0.44. RT also became faster in later blocks, producing a significant main effect of block, F(6, 372) = 11.65, p < 0.001, ηp2 = 0.16. RT was faster in the detection task than the discrimination task, F(1, 62) = 55.94, p < 0.001, ηp2 = 0.47. The lack of interaction between target quadrant and task suggests that probability cuing was comparable between the two tasks, F(1, 62) = 2.10, p = 0.15. Improvement in RT across training blocks was larger in the discrimination task than the detection task, F(6, 372) = 2.17, p = 0.04, ηp2 = 0.03 for the interaction between block and task. None of the other interaction effects were significant, largest F(6, 372) = 1.79, smallest p = 0.10.
Even though participants acquired probability cuing, this effect did not transfer in the testing phase when the task changed. An ANOVA using task, target quadrant, and testing block as factors showed no effects of target quadrant, F < 1. RT improved across blocks, producing a significant main effect of testing block, F(3, 186) = 5.66, p = 0.001, ηp2 = 0.08. RT was faster in the detection task than the discrimination task, F(1, 62) = 38.49, p < 0.001, ηp2 = 0.38 for the main effect of task. Target quadrant did not interact with block, neither did it interact with task, Fs < 1, and the three-way interaction was not significant, F < 1.
To examine the strength of the null effect in relation to the presence of a transfer effect, we conducted a Bayesian analysis on the effect of target quadrant in the testing phase (see “Method”). The Bayesian analysis provides strong evidence that location probability learning did not transfer to the testing phase when the task changed. The Bayes factor of target quadrant was 8.99, suggesting that it was nine times more likely that target quadrant did not affect RT than it did.
The above analysis combined data across all 64 participants who produced behavioral data. Note that half of these were tested on an eye tracker and the other half were not. When “eye-tracking status” was included as a between-group factor in the analysis, this factor did not interact with any of the experimental factors. In the training phase, the interaction between eye-tracking status and target’s quadrant was not significant, F(1, 62) = 1.16, p > 0.28. Location probability learning was significant in each group, F(1, 31) = 20.09, p < 0.001, ηp2 = 0.39 for those with eye-tracking; F(1, 31) = 27.11, p < 0.001, ηp2 = 0.47 for those without eye-tracking. In the testing phase, there was no interaction between eye-tracking status and target’s quadrant, F(1, 62) = 1.13, p > 0.29. Transfer of learning was not significant for either the eye-tracked group, F < 1, or those without eye tracking, F < 1.
Eye movement data: fixation pattern
Training phase
Differences in eye movement provided insight into the lack of transfer between tasks. Both tasks entailed a large number of fixations (Fig. 4, left). Participants made more fixations on target-absent trials than target-present trials. We conducted an ANOVA on the number of fixations, using task as a between-subject factor and target status (target-present or target-absent) as a within-subject factor. This analysis showed a significant main effect of target status, with more fixations on target-absent trials, F(1, 30) = 422.87, p < 0.001, ηp2 = 0.93. Participants performing the discrimination task made more fixations than those performing the detection task, F(1, 30) = 18.06, p < 0.001, ηp2 = 0.38, a difference that was larger on target-present trials than target-absent trials, F(1, 31) = 5.08, p = 0.03, ηp2 = 0.15 for the interaction between task and target status.
Although participants in the discrimination training task made more fixations, on average each fixation was briefer (Fig. 4, right). An ANOVA using target status as a within-subject factor and task as a between-subject factor showed that fixation duration was shorter on target-absent trials than target-present trials, F(1, 30) = 40.74, p < 0.001, ηp2 = 0.58. Fixation duration was longer in the detection task than in the discrimination task, F(1, 30) = 55.34, p < 0.001, ηp2 = 0.65, a difference that was larger on target-absent than target-present trials, F(1, 30) = 5.40, p = 0.03, ηp2 = 0.15 for the interaction between task and target status.
Testing phase
The pattern of fixation data was replicated in the testing phase (Fig. 5). Specifically, participants performing the discrimination task made more fixations than those performing the detection task, particularly on target-present trials, t(30) = 6.59, p < 0.001. Mean fixation duration was briefer in the discrimination task than the detection task, F(1, 30) = 36.86, p < 0.001, ηp2 = 0.55. Other aspects of the statistical analyses were similar to those of the training phase and the details are omitted.
The fixation data showed that both detection and discrimination tasks involved a large number of fixations, supporting the assumption that the tasks required serial search. Differences between the two tasks were also apparent. Participants made more fixations in the discrimination task than in the detection task. However, each fixation was briefer in the discrimination task.
Eye movement data: first-saccadic eye movements
Not only were people faster in finding the target in the high-probability quadrant, but they also acquired a tendency of saccading toward that quadrant first. This effect was most clearly revealed on target-absent trials, where saccade could not have been influenced by the presence of target features (Fig. 6). With four quadrants, the chance rate of saccading toward the high-probability quadrant is 25%. In the training phase (Blocks 1–7), the mean percentage of trials with first saccades to the high-probability quadrant was 49.8% in the detection task and 39.6% in the discrimination task, both of which were significantly higher than 25%, t(15) = 4.97, p < 0.001 for the detection task and t(15) = 2.28, p = 0.04 for the discrimination task. An ANOVA using task and training blocks as factors showed that participants’ tendency to saccade toward the high-probability quadrant increased across training blocks, F(6, 180) = 9.36, p < 0.001, ηp2 = 0.24 for the main effect of block. The main effect of task (F(1, 30) = 1.60, p = 0.21) and the interaction between block and task (F(6, 180) = 1.37, p = 0.23) were not significant.
As the task changed in the testing phase, the saccade pattern also changed. Those trained in the discrimination task and tested in the detection task no longer persisted in their overt search pattern. In these participants, the mean percentage of trials with first saccades to the previously high-probability quadrant in the testing phase was 22.6%, a level not significantly higher than chance, t < 1. Those trained in the detection task and tested in the discrimination task, however, showed a persisting but declining trend of saccading toward the high-probability quadrant. For these participants, the percentage of first saccades directed toward the high-probability quadrant in the testing phase – 41.5% – was significantly higher than chance, t(15) = 2.36, p = 0.03. This effect declined across testing blocks, F(3, 90) = 4.31, p = 0.007, ηp2 = 0.13.
Qualitatively similar results were observed on target-present trials (Fig. 7 in the Appendix). These trials presented some complications given that the first saccades may be made after detecting target features and therefore would be influenced by where the target was on a trial. Nonetheless, the pattern of data was similar to target-absent trials. Specifically, when trained with the discrimination task, participants gradually acquired a tendency to saccade toward the high-probability quadrant first. This preference ceased when the task changed to detection. When trained with the detection task, participants also acquired a tendency to saccade toward the high-probability quadrant first. This preference was substantially reduced, though somewhat persistent, when the task changed to discrimination. Detailed results and statistical analyses that took into account the target’s location can be found in the Appendix.
Figure 7 in the Appendix showed different saccade patterns between the detection and discrimination tasks. Regardless of which phase these tasks were performed in, first saccades in the discrimination task were insensitive to the target’s actual location. The proportion of first saccades toward the high-probability quadrant was no stronger when the target itself was in the high-probability quadrant than when it was elsewhere. This suggests that the first saccades were executed before acquiring target features. In contrast, in the detection task, the proportion of first saccades toward the high-probability quadrant was stronger when the target was in that quadrant than when it was elsewhere. This suggests that first saccades were initiated after participants had analyzed the image and had some information about where the target was. In fact, first saccade latency was longer in the detection task than in the discrimination task, both in the training phase (target-absent trial means: 240 ms vs 156 ms, t(30) = 4.67, p < 0.001 on trials) and in the testing phase (226 ms vs 148 ms, t(30) = 5.50, p < 0.001).
Discussion
Experiment 2 successfully induced a change in spatial attention in the training phase in both the detection and discrimination tasks. However, no transfer in RT was observed when the task changed. This was the case even though the two tasks were performed in the same general space, the task set was similar, and the displays had similar visual characteristics including the use of 1/f3 noise. On its own, the lack of transfer may be explained by differences between the two tasks. For example, the discrimination task took longer. However, differences in search RT did not prevent transfer in previous studies. Jiang, Swallow, et al. (2015) observed transfer between two T-among-L search tasks of different difficulty. The easy task had a mean RT around 1 s and the difficult task 3 s. The discrepancy in task difficulty in that case was greater than in the current study, where RT differed by about 0.5 s. A difference in display appearance (e.g. noise opacity level) also could not explain the results. Salovich et al. (2017) showed transfer between two visually very different tasks – finding a T-among-L and finding an arrow in natural scenes.
What might account for the lack of transfer in the current study? We suggest that the lack of transfer may reflect differences in how search was conducted between the two tasks. The discrimination task requires participants to make serial shifts of attention among items that are easily segmented from the background. The detection task has few candidate regions to inspect but requires longer scrutiny when one is identified. This differs from previous studies where all tasks involve serial scanning among segmented objects. The eye data supported this suggestion. The discrimination task involved a higher number of fixations than the detection task, but each fixation was briefer. In addition, the detection, but not the discrimination, task involved an initial stage of image analysis before the first saccade was made. These data suggest that the search procedures differed between the two tasks.
The target’s location probability not only enhanced search RT, but also induced a tendency to direct the first saccade toward the high-probability quadrant. Consistent with RT, the first-saccade bias acquired in the discrimination task did not transfer to the detection task. However, the saccade preference acquired in the detection task only gradually declined when the task changed to discrimination. This latter finding was not accompanied by an RT advantage. This discrepancy suggests that a habit involving saccades is harder to correct than the covert search habit indexed by RT. The lack of an RT advantage suggests that information gathered from the preferential saccades is discounted at a later level; hence, there was no RT advantage even though eye movements showed a residual preference toward the previously high-probability quadrant.