In the experiments described here, subjects were asked to perform an attentionally demanding visual discrimination task. On some trials, a task-irrelevant, abrupt-onset distractor item would appear prior to the main task array. Initially, the appearance of these distractor items led to a reduction in the accuracy of responses and an increase in response time. After repeated exposure, the distractor-induced RT slowing declined (
Experiment 1). When the appearance and location of the distractors were highly predictable, these practice effects were disrupted by the appearance of distractors in new locations (
Experiment 2) and of new distractor items (
Experiment 3). When the appearance of the distractors was more heterogeneous, subjects required more exposure to the task in order to overcome their distracting effect; however, these practice effects transferred more readily to new distractor conditions (
Experiment 4). Thus, practice with a heterogeneous distractor set led to more generalized improvement in the selection efficiency.
The conclusion that subjects required more practice to overcome distraction with heterogeneous distractors is based on the fact that in
Experiment 1–
Experiment 3, only the first block and the block immediately following a distractor switch (i.e., Blocks 1 and 6) exhibited an RT difference that was significantly greater than zero. In contrast, the RT difference in
Experiment 4 lasted for several blocks. This difference might be due in part to a difference in the magnitude of the initial distraction effect.
Figure 11 shows the RT difference for the first ten trial pairs in each of the above experiments. This plot does show some variability in the initial distraction level between the different experiments. A between-subjects ANOVA showed that Experiment had a significant effect on initial RT difference (
F(3,111) = 3.45,
p < 0.05). However, post-hoc analyses using Tukey's HSD only showed a significant difference between
Experiments 3 and
Experiment 4 (
p < 0.02); no other pairs were significantly different (all
ps > 0.17). As discussed earlier, the subjects in
Experiment 3 were very efficient in overcoming the initial distraction effect. This suggests that the slower reduction in RT cost was not entirely due to the magnitude of the RT difference being greater at the beginning of the experiment.
The pattern of results observed here, that increased distractor variability leads to more effective transfer across distractor conditions, is not consistent with any of the explanations discussed previously. The fact that subjects were able to suppress distractors from multiple classes of objects, all of which were unique images, and transfer this suppression to new locations, strongly argues against sensory adaptation. The lack of transfer in the distractor item transfer condition (
Experiment 3) and the transfer across locations in
Experiment 4 rule out the idea of suppression of specific locations. The quick reduction in the distractor effect for disk/face distractors (
Experiment 1–
Experiment 3) and the slower reduction when there are many classes of distractor item tend to support an object-based suppression mechanism. However, the difference in the levels of transfer to new locations in the different experiments seems to contradict this: more efficient suppression of distractors should in principle produce more effective transfer. This suggests that a different mechanism may be responsible for the reductions in the distractor effect.
On possible explanation for the observed results is habituation of the orienting response (OR), a process that has been extensively studied (Sokolov,
1963; see Bradley,
2009, for a review). The OR consists of various physiological and neural responses to the presence of emotional, salient, or novel stimuli. These responses are typically considered to be indicators that attention has been directed to, or captured by, the stimuli. The OR diminishes quickly with repetition of the stimuli but is reinstated with the presentation of another novel stimulus. This paradigm has typically been used to assess the significance of feature differences between two or more similar objects, similar to adaptation patterns in fMRI (as reviewed by Krekelberg et al.,
2006).
Habituation of the OR may account for the reduced distractor effect observed in
Experiment 1 and
Experiment 2 and, to a lesser extent, in
Experiment 3. In these experiments, a single stimulus (the red and green disks) is presented repeatedly across multiple trials, in relatively stable locations. These circumstances would promote rapid reduction in the OR. If subjects are no longer orienting to the distractor, it should cease to cause a slowing in the response to the main task. The uniformity of the disks then leads to reinstatement of OR following a location change. Elliott and Cowan (
2001) explained a reduction in distraction during a cross-modal Stroop paradigm as due in part to a similar habituation of the OR to auditory distractors.
With the introduction of the more variable distractor sets in
Experiments 3 and
Experiment 4, the current results become less amenable to explanation by appealing to a habituated OR. As discussed previously, Gati and Ben-Shakhar (
1990; as well as Bradley,
2009) present evidence that habituation can be based on repetition of specific features across multiple images, rather than just repetition of a single stimulus. However, the faces used in
Experiment 3 were suboptimal stimuli for habituation to specific features, both in terms of presentation and set heterogeneity. This is doubly true for the distractors used in
Experiment 4. One could argue that the increased time required to reduce the distraction effect in
Experiment 4 represents gradual habituation to multiple feature sets across the full range of stimuli. However, the use of many different stimulus categories makes the number of features that would need to be habituated so large as to be implausible, or so general as to be perceptually untenable (e.g., habituation to curved or straight edges, etc.). This suggests that the observed reductions in the distractor effect, and their transfer to new spatial locations, are the result of improvements in the allocation of attention, rather than habituation of an orienting response. Future studies can further clarify this point by testing transfer of attentional learning effects to previously unseen distractor categories.
That subjects took longer to develop efficient selectivity with heterogeneous distractors suggests a system for distractor suppression that scales with the amount of information available. Where distractors are highly predictable and stable in their appearance, filtering mechanisms can be very narrow and specific to the particular stimuli to be ignored. Such a system would require inhibition of relatively few cortical regions corresponding to either object representation or spatial location, and so should be quickly reinforced over time, reaching peak efficiency with only a few repetitions. Additionally, the cognitive demands of such a filtering mechanism would be relatively low, making it an effective strategy for performing a task such as the one used here.
Where the distractors are highly variable and unpredictable, a stimulus-specific filtering mechanism would prove ineffective. Instead, a more general filtering mechanism would be required, one that could be applied in many different conditions. Because the first few distractor exposures would be similar to a predictable distractor condition (with limited sampling of the distractor set, the level of heterogeneity in the distractor set would be indeterminate), the filtering system would be established in a bottom-up fashion, updating as more information became available. Such a system would be slow to reach peak efficiency for two reasons. First, as new distractor items continued to appear, they would be outside the bounds of the system currently in place, and so would not be effectively filtered. This would initially yield only minimal reduction in the distractor effect until a sufficiently broad filter was established. Second, a general filtering mechanism would still require enough constraints that it did not impair perception of task-relevant information (e.g., the array of red and green dots). Thus, inhibiting all sensory input in a given modality could not be used. It follows then that a filtering mechanism that would work within the constraints of the current task would involve inhibitory interactions among a variety of visual processing areas, such as retinotopically mapped early visual cortex and object-selective inferior temporal cortex. Because these inhibitory connections would cover many functionally specialized regions, making the distractor filtering system more complex overall, any single distractor item would necessarily provide less information for the execution of the filtering system, compared to the case with highly predictable distractor items. Thus, less variable distractors would be overcome more quickly because the system for filtering out the distractor information would be more specific, and thus implemented more quickly. In contrast, a more general filtering mechanism would take longer to be established.
Such a mechanism, in which more generalized training leads to greater transfer to new distractor conditions, is supported by work in both verbal memory and motor learning, as reviewed by Schmidt and Bjork (
1992). Here, the authors describe several studies investigating many different tasks showing that conditions that promote rapid improvement during practice do not lead to optimal post-training performance. Rather, training conditions that typically lead to worse performance during practice produce more complete learning and improved performance in a wider variety of post-training scenarios. The authors argue that this is because the more difficult and varied practice conditions lead to deeper and more complete information processing, necessarily leading to better performance. The present findings, and proposed models, are consistent with this idea, drawing an important parallel between attentional learning and skill training in general. This has important implications for future studies of attentional processes.
The results discussed here diverge from previous examinations of improvements in attentionally demanding tasks with repeated exposure. The experiments of Schneider and Shiffrin (
1977) showed improvements in target detection when the set of target items was repeated over many trials; perturbations in this set eliminated all practice effects. The contextual cueing paradigm (Chun & Jiang,
1998,
1999) leads to improved task performance because the distractors actually guide attention, rather than being filtered (though see Makovski et al.,
2008, for an example of learning restricted to the target items). Brown and Fera (
1994) showed a reduction in the flanker effect with practice, but only under limited circumstances and with distractors that carried task-relevant information (the flankers were predictive of the main target); these effects also seemed to be highly dependent on the particular parameters and timing of task, suggesting limited possibility for transfer. These studies are examples of the general trend in tasks that examine practice effects and attention: improvements are tied to the particular task-relevant stimuli and generally involve the development of automatic associations between stimuli and the correct response.
A recent study by Dixon, Ruppel, Pratt, and De Rosa (
2009) also examined the issue of improvement in the allocation of attention and filtering of distracting information. In their study, subjects were asked to identify which of a pair of objects belonged to a previously defined target set based on color and shape. During priming blocks, only color information was useful for identifying the target item, with the shape of the objects being non-informative. During the probe blocks, however, subjects needed to use the shape of the objects to identify the targets, and color became non-informative. Subjects were slower to respond to the shapes in the probe blocks after having ignored them in the prime blocks; this effect was largest when the shapes were previously ignored in two separate color contexts, indicating generalization of the filtering process. The findings of Dixon et al. are mirrored by the current results, which examine improved filtering not for features of attended objects but for completely separate distractor items. Of particular importance is the more generalized learning effect produced by a more variable training condition found in both studies, again corresponding to the processes described in other skill training by Schmidt and Bjork (
1992).
As discussed earlier, there is some evidence for improvements in the performance of working memory tasks with practice, though it is not consistent. Where such evidence does exist (Jaeggi et al.,
2008; Klingberg et al.,
2002; Olesen et al.,
2004), it seems to suggest that effective training on a single working memory task leads to general improvements in working memory capacity and overall cognitive ability. Because the current studies only investigated transfer across different distractor conditions, it remains to be seen whether improvements in distractor filtering lead to general improvements in selective attention in different task domains. As such, the relationship between the current results and improvements in working memory performance remains uncertain.
The improvements in distractor filtering observed here raise the possibility that perceptual learning may have been a factor. Studies of perceptual learning involve extensive practice with identification and discrimination of subtly different visual stimuli (for reviews, see Ahissar & Hochstein,
2004; Fine & Jacobs,
2002; Petrov, Dosher, & Lu,
2005). In these studies, the extent to which learning transfers to new conditions is highly dependent on the stimulus and task parameters. In their Reverse Hierarchy model, Ahissar and Hochstein (
2004) have suggested that transfer of training is dependent on the perceptual processing level at which the learning took place. According to this account, greater transfer would be expected where relatively coarse discriminations can be accomplished by later processing stages (e.g., discrimination between objects, relying on IT cortex). However, more difficult discrimination requiring more fine-tuned analysis would produce less transfer because it relied on changes in a more narrowly responsive set of neurons (e.g., a single orientation column in area V1). Alternately, Dosher and Lu (
1999,
2007; Petrov et al.,
2005) argue that learning is the result of fine-tuning of perceptual templates, resulting in improved filtering of external noise. They further argue that such fine-tuning only happens under difficult discrimination conditions (i.e., with confusable targets in the presence of distracting information), and that because it manifests as a change in the strength of connections between perceptual and response processing stages, it is specific to the learned stimulus and the context in which it was viewed.
At first glance, the present data seem to contradict these models, given the greater transfer observed in
Experiment 4 compared to
Experiment 1–
Experiment 3. However, several aspects of the current paradigm make perceptual learning models a poor fit for these results. Though the distractors, being complex objects, would have been processed in object responsive regions of visual cortex (where the Reverse Hierarchy model states that generalized learning is more likely), there was no need for improved discrimination of these items. Their locations were predictable, and none of them resembled the target array, making the improved discrimination afforded by perceptual learning unnecessary for any improved filtering of the items. Furthermore, the distractors were spatially distinct from the dot array and several degrees into the periphery, reducing the degree to which they could visually interfere with the array. Lu, Lesmes, and Dosher (
2002) showed that endogenous attention improved perceptual template mapping in much the same way as perceptual learning, by reducing the effects of external sensory noise. They also showed that this reduction in external noise is confined to the target location and seems to have little impact on adjacent locations when the target location is known in advance (as was the case here). This suggests that perceptual learning accounts are not adequate to explain the present results.
Though the proposed distractor filtering mechanism that develops based on information about the distractor set addresses how improvements in performance arise, it does not address the nature of those improvements. How does practice change the way in which the distractors are processed so that they no longer slow responses? The distracting items used here were all abrupt onsets, a class of stimuli that have been shown to effectively capture visuospatial attention (see Yantis & Jonides,
1996); additionally, the distractors share certain features with the task array (the red and green disk distractors have the same colors as the array; both distractors and the task array appear abruptly on the display) that could facilitate contingent capture based on subjects' attention priority sets (see, e.g., Folk & Remington,
1998). Because the distracting items successfully capture attention, they consume limited attentional resources, which are necessary for task performance, producing slower response times. Previous work on attentional capture, though, does not indicate that subjects become less susceptible to capture over time. This suggests that the reduction in the distractor effect may not be the result of a reduction in attentional capture.
A key difference between this paradigm and others is that most other experiments involving attentional capture show decrements to performance in situations of spatial uncertainty. For example, capture in a visual search task usually involves a stimulus that draws attention to its spatial location. In situations when this stimulus is not the target item, subjects must then continue searching for the target in the rest of the display. The capture of attention in this case has produced a delay in the response (see also Folk et al.,
2002, for an example of capture in a situation of temporal uncertainty). In the current task, there is neither spatial nor temporal uncertainty. The task array always appears in the same location and at a very regular interval from one trial to the next. This certainty could allow subjects to adapt to the presence of the distracting items and eventually avoid being captured by them. If temporal and spatial regularities are important factors in overcoming attentional capture, then it follows that more regular distractors would be easier to overcome, as was the case here. In contrast, distractors that had a more variable appearance might require a more general filtering mechanism to avoid capture by task-irrelevant stimuli, accounting for the greater transfer of practice effects in those cases.
Alternately, it is possible that subjects are not overcoming the capture of attention, but rather the tendency to then process the identity and behavioral relevance of the distractors. In this case, the sensory information about the distractors might still be processed, but that information would no longer interfere with the cognitive operations of the main task. Such an improvement would also be tied to the level of variability in the distractor items, resulting from either the rapid adaptation of an information filtering mechanism that was specific to the observed distractors, or a more general one that could be applied to previously unobserved conditions.
These two explanations of the observed improvement in performance lead to two different predictions about how activity in the brain changes along with practice. If the distractors no longer capture subjects' visuospatial attention, then one would expect to see reduced activity in the early visual areas associated with the appearance of the distractors. In contrast, if practice is leading to reduced cognitive interference, then one should observe changes in the activity of more frontal regions, while the response in early visual areas remained stable. Such predictions could be addressed directly by adapting this paradigm for functional neuroimaging.
Selecting task-relevant and ignoring task-irrelevant information is a critical aspect of successful goal-directed behavior. The experiments described here addressed these issues by examining distraction effects during a visual discrimination task. The data revealed that repeated exposure to distracting stimuli led to an improved ability to filter out the distracting information. Furthermore, exposure to a more variable set of distractors produced more effective filtering that could be applied to previously unseen locations. These studies show that selective attention improves with training, and that these improvements occur at the cognitive, rather than the sensory, level. This relates improvements in attention to previous studies of skill training in general, which has important implications for the future study of attention and information selection.