The Brain's Sensitivity to Real-world Statistical Regularity Does Not Require Full Attention Free

Over time, as we navigate life, we abstract patterns, or real-world statistical regularities, from the environment. Hierarchical predictive coding models (Clark, 2013; Friston, 2008; Spratling, 2008; Friston, 2005; Rao & Ballard, 1999) view perception as actively predictive and imply that these real-world statistical regularities are encoded within multiple stages of the brain's visual hierarchy and thereby constitute the fabric that makes perceptual predictions possible. What is not well understood is how real-world statistical regularities affect ongoing perception. Do these encodings play a constant, automatic role in basic perceptual processes, or is it the case that active top–down modulation—that is, attention—is required to bring them online? In the next section, we unpack the concept of real-world statistical regularity before relating it to the concept of attention.

Predictive coding models stand in contrast to traditional serial models of perception (Driver & Baylis, 1996; Nakayama, He, & Shimojo, 1995; Palmer & Rock, 1994; Wertheimer, 1923; Rubin, 1915), which tend to view visual processing and conceptual processing as separate stages along a linear trajectory. Challenges to this view (e.g., Grill-Spector & Kanwisher, 2005; Peterson & Gibson, 1994) are bolstered by findings from psychology and neuroscience demonstrating that factors related to a stimulus' identity, such as its canonicity, familiarity, probability, or representativeness, can influence basic perceptual processes (Yang & Beck, 2023; Center, Gephart, Yang, & Beck, 2022; Kumar, Federmeier, & Beck, 2021; Smith & Loschky, 2019; Caddigan, Choo, Fei-Fei, & Beck, 2017; Greene, Botros, Beck, & Fei-Fei, 2015; Torralbo et al., 2013). Each of these factors contributes to what we have termed “real-world statistical regularities” (Shao & Beck, 2024; Yang & Beck, 2023; Center et al., 2022)—that is, patterns extracted from our visual environment that serve to minimize prediction error with respect to determining what is being viewed. In other words, these regularities are patterns that will allow, on average, the visual system to more efficiently perceive and recognize what is out there in the world.

For instance, Caddigan et al. (2017) showed participants good and bad exemplars of natural scenes (the same scenes used here) at very brief presentations (<78 msec) and asked participants to report whether the scenes were intact or scrambled (i.e., phase-scrambled noise companions of intact scenes). Despite the facts that participants were given no information about the category labels and that image category was irrelevant to the task, researchers still observed an effect of real-world statistical regularity on detection, with the result that good exemplars (i.e., those that had been rated as more representative of their category) were better discriminated from noise than bad exemplars. Note that the task was not to identify the scene contents but merely to indicate whether the stimulus shown was a real scene or just random noise. This detail makes a strong case for the fact that even early perceptual processes are affected by “higher-level” regularities yoked (in this case) to scene categories. Predictive coding models posit that predictions are generated at each level of the visual processing hierarchy and that access to perceptual predictions derived from statistical regularities renders basic perceptual processes more efficient. In other words, visual processing is faster and more effective when stimuli match perceptual predictions, which, in turn, have been shaped by the goal of ultimately recognizing scenes and objects, such that the statistical regularities that matter for processing, even at early levels, are those that matter for identifying and categorizing visual inputs. It is unclear, however, whether the activation of said regularities and subsequent matching processes require attention.

The predictive coding literature draws a clear distinction between the prediction (or, as it is sometimes called, expectation) and attention (Feldman & Friston, 2010; Friston, 2009). The prediction in this sense is an implicit expectation internalized within the network, or a prior in a Bayesian framework, that is informed based on statistical regularities encountered in the world that have previously proven useful for recognition. Attention, on the other hand, is the process of prioritizing certain features or spatial locations on the basis of task relevance. Several lines of research support the claim that expectation and attention reflect different mechanisms (Rungratsameetaweemana & Serences, 2019; Summerfield & Egner, 2016; Kok, Rahnev, Jehee, Lau, & De Lange, 2012; Summerfield & Egner, 2009). Thus, from the predictive coding perspective, there is no reason to believe that generating useful perceptual predictions should require attention (although it may be modulated by attention; see Feldman & Friston, 2010). Furthermore, in behavioral experiments using natural scene stimuli (of the same kind used in this experiment), there is evidence that extracting the gist of a scene requires little to no attention (Fei-Fei, VanRullen, Koch, & Perona, 2005; Li, VanRullen, Koch, & Perona, 2002; Rousselet, Fabre-Thorpe, & Thorpe, 2002; Thorpe, Fize, & Marlot, 1996; Potter, 1976). Assuming that the processes involved in extracting gist-based information also involve evaluating perceptual predictions on the basis of real-world statistical regularities, then it would seem that attention is not necessary for predictive coding-based processing of scenes.

One might, nonetheless, expect that attention could play a role in bringing appropriate perceptual predictions online. Some have argued that attention is in fact necessary to extract gist, at least when the task is sufficiently difficult and attentional demands of a distracting task are sufficiently high (Mack & Clarke, 2012; Cohen, Alvarez, & Nakayama, 2011). Similarly, Lavie, Lin, Zokaei, and Thoma (2009) argued that participants fail entirely to form representations for distractors under conditions of high perceptual load, wherein attention is stretched thin. The role of attention has also been probed in studies of statistical learning, although (as discussed more below), the kinds of real-world statistical regularities we are investigating are acquired over long time periods, and thus effects observed in statistical learning paradigms may differ, because they typically tap into patterns acquired within the context of an experiment. In the context of statistical learning studies, it has been argued that visual statistical regularities bias the allocation of attention (Wang, Samara, & Theeuwes, 2019; Wang & Theeuwes, 2018a, 2018b; Baker, Olson, & Behrmann, 2004) and might even require attention at some level to obtain (File & Czigler, 2019; Turk-Browne, Jungé, & Scholl, 2005; but see also Duncan & Theeuwes, 2020). In the auditory domain as well, Toro, Sinnett, and Soto-Faraco (2005) made the case that attention is needed to enable statistical learning of speech segmentation of nonsense language streams (but see also Batterink & Paller, 2019). Finally, some research using fMRI has argued that attention is necessary for, or is interactive with, expectation (Richter & de Lange, 2019; Kok et al., 2012; Larsson & Smith, 2012), whereas more recent work using a design that leverages real-world statistical regularities has shown that full attention is not necessary for the good exemplar advantage, although aspects of the advantage may be modulated by attention (Shao & Beck, 2024). We note, however, that because of the low temporal resolution of fMRI, these results likely conflate early and late advantages of representative scenes categories. Thus, it remains unclear whether the impact of real-world statistical regularity on early processing and the N300 might also require attention.

Although the role of attention in the use of learned regularities has been probed more extensively in studies of statistical learning (Sherman, Graves, & Turk-Browne, 2020; Saffran, Aslin, & Newport, 1996), where statistical regularities are typically directly manipulated within the trial structure of an experiment, there are reasons to question whether such findings can be generalized to the kind of regularities we are studying in the present work. We refer to the regularities we are focusing on as “real-world” statistical regularities because they are built up over the course of a lifetime rather than over the course of an experiment. Although there are a growing number of exceptions (e.g., Forest, Schlichting, Duncan, & Finn, 2023; Sherman et al., 2020; Smith, Jayaraman, Clerkin, & Yu, 2018; Kóbor, Janacsek, Takács, & Nemeth, 2017; McCauley, Isbilen, & Christiansen, 2017; Brady & Oliva, 2008), the majority of statistical learning research has looked at sensitivity to patterns acquired over the (obviously much shorter) time course of an experiment and has also tended to use stimuli consisting of simple shapes or sounds and to focus on manipulations of transitional probabilities of repeated stimuli (Frost, Armstrong, & Christiansen, 2019). This approach provides tight control over the statistical regularities that a participant is experiencing, but differs in important ways from regularities acquired from more complex stimuli, such as natural scenes, over much longer time courses.

Varying stimuli on the basis of real-world statistical regularities takes advantage of previous experience, rather than treating it as a nuisance variable, and in doing so allows researchers to study more complex types of stimuli and patterns of regularities over them that arise in our daily experience. Although some shared neural mechanisms likely support learning in both scenarios, beyond clear differences in time scale and stimulus complexity, looking at real-world statistical regularities shifts the research focus from local co-occurrence relations among repeated stimuli to meaningful variation of unique stimuli within a particular class. In this sense, our concept of real-world regularities shares more in common with the category learning literature. Furthermore, the outcome measure of interest is not what has been learned (assessed via poststimulus memory tests and/or RT measures), but what the person perceives in the moment as a function of prior learning (assessed via stimulus-concurrent neural responses and/or detection tasks). In summary, our research shares some commonalities with the concepts studied in statistical learning; indeed, real-world regularities would likely be acquired via some form of statistical learning. In practice, however, the regularities typically studied in the statistical learning literature tend to differ in important ways from the ones studied here, and the highlighted differences are sufficiently critical to warrant the study of real-world statistical regularity in its own framing.

ERPs offer temporally specific and functionally well-characterized measures of brain processes, which can help us to draw key inferences about the nature and use of implicit and explicit predictions. One component that may be particularly useful for addressing questions about these real-world statistical regularities acquired implicitly over the long-term is the N300, a fronto-centrally distributed negativity that is observed during the visual processing of objects and scenes (Schendan, 2019; Schendan & Kutas, 2003; Schendan & Kutas, 2002). The N300 typically peaks around 300 msec, later than sensory components (e.g., the C1, P1, and N1) that are sensitive to low-level visual features (see Luck, 2014, chap. 3), but before multimodal responses like the N400, which has been taken to index access to long-term semantic memory (see Kutas & Federmeier, 2011). The N300 is thus well-positioned in time to serve as an index of the use of long-term knowledge in the analysis of incoming visual information.

We have made the case that the N300 is larger for items that are less statistically regular (Kumar et al., 2021). Early experiments elicited N300 differences by displaying line drawings of objects that varied in completeness, and the component was originally viewed as a neural index of the structural integrity of an image (Schendan & Kutas, 2002). Larger N300 amplitudes, however, have also been observed in response to photos of real objects in noncanonical viewpoints, relative to those elicited by objects in canonical viewpoints (Schendan & Kutas, 2003). Take, for example, a coffee cup viewed from the bottom as opposed to the same cup viewed from its side profile. In this case, a larger N300 elicited by seeing the cup from the bottom cannot be explained by structural integrity, as it is still a fully intact object, and likewise, it cannot be fully explained by familiarity (such as in Yang & Beck, 2023 or Manahova, Mostert, Kok, Schoffelen, & De Lange, 2018), as we often see objects from many angles throughout daily life (for instance, when washing a coffee cup; Center et al., 2022). Similarly, a greater N300 is evoked by otherwise typical objects that are inconsistent or improbable with respect to their context (Võ & Wolfe, 2013). These findings suggest a wider role for neural processes tapped by the N300.

Recent experiments have since shown that N300 differences can also be elicited by contrasting highly representative (good) and less representative (bad) exemplars of natural scene categories (Kumar et al., 2021). An example of a good exemplar in this case might be the kind of beach you would see on a postcard at a popular tourist destination (e.g., one with blue sky, white sand, and clear water), whereas a bad exemplar might be a beach that satisfies the criteria for the beach category but contains less canonical variants of beach features (e.g., one with overcast skies, a littered shore, and/or murky water). Exemplars in these experiments came from six natural scene categories. Good exemplars were rated by independent observers as highly representative of their category, whereas bad exemplars, although deemed exemplars of their category, received low ratings of representativeness (Torralbo et al., 2013). Although the real-world statistical regularity of scenes in previous studies (Kumar et al., 2021; Caddigan et al., 2017; Torralbo et al., 2013; and also the current experiment) was operationalized via the aforementioned representativeness ratings, we believe many factors may contribute. Some good exemplars might be more canonical, with more informative or diagnostic features or properties present than in bad exemplars. Some good exemplars might be more familiar; although it is unlikely participants have previously viewed the photos of the exact scenes used in these studies, and they are only shown once in the experiment, images consumed online tend to be good exemplars of their categories. Examples of our good and bad natural scene exemplars are shown in Figure 1.

Across multiple previous studies (Smith & Federmeier, 2024; Kumar et al., 2021; Smith & Federmeier, 2020; Võ & Wolfe, 2013; Schendan & Kutas, 2003; Schendan & Kutas, 2002), less expected or less predictable stimuli consistently induce greater N300 amplitudes. We, therefore, have interpreted this pattern as supporting a broad sensitivity of the N300 to variations in real-world statistical regularity—that is, the degree to which the stimulus matches an implicit expectation or prediction, thus requiring less intensive processing to resolve its identity (Kumar et al., 2021). This interpretation is similar to that of Schendan's object model selection view (Schendan, 2019; Schendan & Kutas, 2002, 2003, 2007), in which the N300 reflects the process by which the best match to the percept, among other candidate objects, is selected. Our interpretation extends beyond objects, however, and we emphasize the predictive aspect of this process. The visual system is making a prediction about the percept, and those predictions reflect real-world statistical regularities constructed over a lifetime of meaningful interaction with the environment. As attention was never manipulated in previous experiments examining the N300, it is yet unknown whether attention is required to engage these predictions.

In reflecting on the role of attention in assessing real-world statistical regularity, it is useful to consider the temporal relationship of the N300 relative to two other ERP components that have been linked to prediction, but show differential effects of attention. These are the (visual) mismatch negativity (MMN and vMMN) and the P3b. Studies examining the vMMN reveal that, in a slightly earlier time window than that of the N300, the brain quickly becomes sensitive to disruptions of recent sensory patterns even when participants are not paying attention (for a review, see Kimura, 2012). On the other hand, studies using the P3b have shown that other brain systems are also sensitive to pattern disruptions in the local context, but in a manner that does depend on attention (for a review, see Polich, 2007), and this component generally occurs in a slightly later time window than that of the N300. Noting that the N300 is sandwiched in latency between these two components that differ in their sensitivity to manipulations of attention highlights the question of whether the N300 itself is sensitive to attention. The N300 also differs from both the MMN and P3b in that both of these effects have been more closely associated with the processing of short-term statistical patterns dictated by the design of the experiment, rather than those naturally acquired over the long term in the real world, as in the case of the N300.

Are N300 differences then elicited automatically, similar to the MMN, or is attention necessary to mediate feedback from long-term memory structures, similar to hypotheses about the role of attention in mediating the processing of patterns in working memory, as indexed by the P3b? Source localization of the N300 elicited by objects indicates generators in ventrolateral prefrontal cortex and late visual areas of occipitotemporal cortex (Allison, Puce, Spencer, & McCarthy, 1999; Puce, Allison, & McCarthy, 1999), and these areas are known to exhibit sensitivity to top–down modulation by attention (for a review, see Moore & Zirnsak, 2017; Kastner, McMains, & Beck, 2009). It should also be noted that the time course of the N300, peaking near 300 msec, provides plenty of time for feedback processes to occur (see Foxe & Simpson, 2002, e.g., who argue that even some activity in early visual areas occurring pre-100 msec poststimulus onset reflects feedback). Putting these pieces together renders plausible the idea that attention might be required to elicit N300 effects.

We take aim at the question of whether attention is required for the assessment of one type of real-world statistical regularity by comparing N300 amplitudes elicited by exemplars of natural scenes that vary in their representativeness, while participants are either attending (or at minimum, passively viewing) the scenes, similar to previous work (Kumar et al., 2021). The resulting activity is compared with that from a paradigm wherein participants must divert their focus toward an attentionally demanding task while task-irrelevant scenes are presented in the background. In a recent complementary fMRI study, Shao and Beck (2024) used a challenging rapid serial visual presentation task to manipulate attention while participants were presented with a subset of the natural scene categories (cities and mountains) used here. The researchers found that activations in the scene-selective parahippocampal place area (PPA) were stronger to bad exemplars than good, and that although attention further boosted this signal, attention was not necessary to produce the differentiation between good and bad exemplars in PPA. Although this fMRI study can tell us how attention and representativeness interact in specific brain regions, the nature of the BOLD signal does not allow us to disambiguate processes that precede semantic access from those that follow (Shao & Beck, 2024). In contrast, the current study uses the time-sensitive method of ERPs that allows us to look at the brain's response to attention and representativeness in a latency window that precedes semantic access (i.e., before the N400 response). Such data are necessary to assess the hypothesis that attention is not necessary for the brain to engage in visual predictions.

We also have a secondary goal, which is to ask whether short-term statistical regularity, that is, patterns that can be acquired within the context of an experiment similar to those manipulated in statistical learning paradigms, impacts the N300. Indeed, previous research using repeated simple stimuli have shown that statistical learning may occur outside attention (Duncan & Theeuwes, 2020; Batterink & Paller, 2019; Turk-Browne et al., 2005). Could participants abstract patterns from sequences of complex, nonrepeated stimuli determined within the experiment, and might that impact the processing of real-world statistical regularities acquired before the experiment, as measured by the N300? Kumar et al. (2021) found that cueing a scene category before presenting a scene exemplar can influence N300 effect amplitudes, indicating that some degree of added artificial structure might affect the way the brain processes real-world statistical regularity. Presenting images randomly without cues forces the brain to rely on a baseline set of predictions based on the image itself; however, when a particular cue is provided, these baseline predictions are seemingly overridden in favor of those related to the cue. Specifically, when the cue was followed by good or bad exemplar of the cue's category, a robust N300 difference was elicited by the good versus bad exemplars, but when the cue mismatched the category of the presented scene (i.e., the word “forest” preceding an image of a beach), the difference between the N300s evoked by good and bad exemplars was eliminated. In other words, when neither the good nor the bad exemplar matched the prediction induced by the prime, there was no longer an N300 difference, suggesting that the difference was indeed related to prediction. In a similar vein, Moore, Robinson, and Mattingley (2024) found that perceptual predictions sharpen the representational fidelity of expected stimuli, leading to a boost in multivariate pattern analysis classification accuracy of ERP responses for those expected objects as well as a decrement in classification accuracy for unexpected objects. However, in these experiments, visual stimuli were processed under conditions of full attention, thus leaving open to speculation the role of attention in those effects.

Therefore, in an exploratory analysis, we follow up on the priming effect found by Kumar et al. (2021) by probing the extent to which artificial regularities alter processing of real-world regularities and whether this manipulation of expectation is modulated by attention. To this end, we manipulated the frequency of good and bad exemplars within each block by using either (1) a completely randomized trial structure in one condition (50–50 good/bad split), which we refer to as the “real-world statistical regularity” condition because any effect of statistical regularity could only be imparted by patterns experienced in the real world or (2) a trial structure blocked by natural scene category in which either good or bad exemplars dominate (80–20 split) within a single scene category, which we refer to as the “artificial statistical regularity” condition because now the experimental context imparts a regularity in addition to the real-world statistical regularities embedded within the images (i.e., the image is more or less likely to be a good or bad exemplar of a single natural scene category). This manipulation is more akin to the trial structure manipulations commonly employed by statistical learning paradigms.

We initially conducted a pilot experiment on only the distracted condition in which participants were asked to play an attentionally demanding game at fixation while task-irrelevant scenes were displayed in the background. Participants were never asked any questions regarding the scenes. The pilot results replicated our group's prior finding of a larger N300 in response to bad compared with good exemplars and further extended this effect to conditions of distraction (for more information, see https://osf.io/kwszv). In particular, the pilot results showed an increased N300 response to bad compared with good exemplars, even when participants' attention was directed elsewhere. The result indicates that full attention may not be necessary to assess representativeness in scene structure. Moreover, there were indications that the N300 effect may be larger under conditions in which real-world statistical regularity (exemplar representativeness) is combined with artificial statistical regularity (local context enacted via increased predictability within the experiment).

The current experiment was designed to replicate and expand upon the pilot experiment by adding an attention condition to directly compare with the distraction condition in a within-subject design. We took the position that prediction/expectation and attention were distinct processes. Thus, our primary prediction was that we would observe greater fronto-central scalp negativity elicited by poorly representative (i.e., “bad”) exemplars than highly representative (i.e., “good”) exemplars, even when stimuli are viewed under distracted conditions (project preregistered at https://osf.io/kwszv). Furthermore, we predicted that the N300 effects in attended and distracted conditions would not meaningfully differ in size (smallest effect size of interest: dz = .3). We were also interested in exploring whether artificial statistical regularity (i.e., the frequency of good and bad scenes within a block) would change the size of the good/bad effect and whether diverting one's attention impacts the brain's sensitivity to this manipulation. We present results from the full data set (n = 34); however, results from only the smaller sample (n = 20) who passed the memory quiz criterion (described below), and our detailed reasons for doubting the efficacy of this exclusion criterion, are fully articulated in the Appendix.

The estimated effect sizes from pilot data (n = 12) under distracted conditions for real-world statistical regularity, artificial statistical regularity, and the difference between the two conditions were Cohen's dz s of 0.79, 1.48, and 0.87, respectively, following the analysis protocol described below. Power analysis indicated that 20 participants were needed to achieve 95% power to detect the smallest effect of interest from the pilot data set (dz = 0.79, one-tailed within-participant t test, α = .05). We recruited 34 participants via the university online paid participant pool to reach 20 participants that passed our original exclusion criteria, described below. Participants were compensated $10 per hour for participation. All participants had normal or corrected-to-normal vision (including no color blindness) and no history of psychiatric disorder or traumatic brain injury. They gave written informed consent in accordance with procedures and protocols approved by the University of Illinois institutional review board.

Stimuli consisted of 720 good and bad exemplars (360 each) from six categories of natural scenes (beaches, cities, forests, highways, mountains, and offices; see Figure 1). Images were 800 × 600 pixels in size and subtended 14.06° of visual angle horizontally and 10.54° of visual angle vertically. Images were presented over a medium gray background. Good exemplars were previously rated by a large sample of independent observers as highly representative of their category, whereas bad exemplars were rated as belonging to the category but poorly representative of it (for full details, see Torralbo et al., 2013).

Participants were seated comfortably approximately 90 cm from a CRT monitor with a refresh rate of 60 Hz. We employed a 2 × 2 × 2 experiment design that manipulated factors of representativeness (good vs. bad exemplars), attention (attended vs. distracted conditions), and context (real-world vs. artificial statistical regularity blocks) as depicted in Table 1. Given that prior work had already characterized N300 good/bad effects under conditions of full attention, the distraction condition was the important one for this design and was set up to ensure that participants could not give their full attention to processing scenes because of the need for them to maintain performance on the “bug task” (described below). Although we expected N300 effects to emerge under conditions of passive viewing with attention, we encouraged participants to keep their eyes on the screen during the attended condition by telling them that their memory for the scenes would be tested. Real-world statistical regularity conditions consisted of six blocks of 60 trials each and artificial statistical regularity conditions consisted of 12 blocks of 60 trials each. Real-world statistical regularity conditions (in which good and bad exemplars were equally probable) always preceded all artificial statistical regularity conditions (in which blocks were either 80% good exemplars or 80% bad exemplars from the same natural scene category), so that effects of artificial statistical regularity would not impact our measure of the effect of real-world statistical regularity. Each trial presented a centrally located exemplar for 200 msec followed by a 500-msec ISI.

The experiment was implemented using Python 2.7 and the open source PyGame module (https://www.pygame.org). Both initial distracted and attended real-world statistical regularity blocks were preceded by a short practice block composed of separate images from those used in main blocks. In the scene-distracted task, nicknamed the “bug game,” participants were asked to continuously monitor a small randomly moving dot (the “bug,” subtending 0.035° of visual angle) and use the control stick to keep it inside a slightly larger circle (the “bubble,” subtending 0.35° of visual angle) in the center of the screen. On every frame, the bug would randomly update its position by −1, 0, or 1 pixels from its previous position in both the x and y coordinates, giving the perception that the bug was rapidly, randomly fluttering from its initial starting position at the center of the screen. In addition, the bug randomly changed color between red and green on 0.005% of frames (resulting in a mean change rate of 0.15 Hz), and participants were required to continuously hold down one of two buttons on a videogame controller corresponding to the bug's color. If the correct button was continuously held, then the participant could override the movement of the bug by 1 pixel per frame on both axes using the joystick, but otherwise the joystick had no influence on the bug's movement. The participants' goal in the distracted blocks was thus to hold down the button that matched the color of the bug at any given time and simultaneously use the joystick to keep the bug inside the bubble. The constant fluttering movement of the bug meant that the participant was required to hold down the proper button and move the joystick nearly constantly to succeed at their task. Meanwhile, in the background, natural scene images with the same timing and size parameters as those in the scene-attended task were presented. Participants were told that scenes would appear in the background but that they were irrelevant to their task. Participants continuously engaged in keeping the bug within the bubble at fixation throughout the block while scenes were presented in the background.

In the scene-attended task, participants were asked to keep their eyes fixated on the center of the screen where the same bug and bubble stimuli ran their course outside of user control. The bug's behavior was identical to the scene-distracted task with the exception that the bug remained within the limits of the bubble without user influence. Meanwhile, participants were instructed that they should pay attention to the scenes as they are presented, while still keeping their gaze fixated on the center of the display. To ensure that participants attended the images, participants were instructed before the start of any scene-attended block that we would test their memory for the scenes at the end of the block. Specifically, each block concluded with a binary forced-choice memory quiz where participants viewed two scenes side by side on the screen and chose which of the two scenes occurred in the previous block. One scene always originated from the previous block, and one scene always originated from a practice block. The position (left or right) of the correct scene was randomized. Participants were asked to press a corresponding button on a game controller to indicate which of the two scenes was presented in the immediately preceding block. For single category blocks (artificial statistical regularity conditions; see below), both scenes always matched the scene category of the preceding block.

In real-world statistical regularity conditions (so named because any “expectation” can be thought to have been built up over a lifetime rather than within the context of the experiment), we assessed the N300 as a function of category representativeness (good vs. bad exemplars). Scene categories were presented randomly and in equal frequencies of good and bad exemplars in each block. To avoid any other forms of predictability (e.g., experiment dictated predictability) besides representativeness, images were never repeated in the real-world statistical regularity condition and participants always completed these conditions before any of the artificial statistical regularity conditions, although the order of attended and distracted conditions was counterbalanced within the real-world and artificial conditions across participants. Artificial statistical regularity conditions (so named because, in addition to real-world factors as above, “expectation” can be built within the artificial context of the experiment) were composed of blocked scene categories and manipulated the frequency of good and bad exemplars, such that good majority blocks had 80% good (20% bad) and bad majority blocks had 80% bad (20% good). For example, a participant may receive a beach block (or one of the other five natural scene categories) consisting of 80% good beaches and 20% bad beaches (or vice versa). Because repetitions might only bolster artificial statistical regularities, the artificial conditions repeated images from the real-world statistical regularity blocks and always followed real-world blocks, although again the order of attended and distracted conditions was counterbalanced within the artificial conditions across participants.

To probe the extent to which the distraction task depleted attentional resources, we included a dual-task oddball paradigm in four final blocks. Each block contained repeated presentations of a single standard image (80% of trials) and a single oddball image (20% of trials) randomly selected from an independent pool of similar natural scene images that have not been rated for representativeness. This procedure was identical to a standard visual oddball task (e.g., a block composed of two repeated stimuli where a majority of presentations are a vertical gabor patch and a minority are a horizontal gabor patch), but used arbitrary natural scenes rather than more simple stimuli. Presentation duration and ISI remained the same as in the main experiment. In the scene-attended condition, participants simply counted the number of oddball scene presentations in each block, whereas in scene-distracted conditions, participants were assigned the additional task of playing the bug game while also counting the number of oddball scene presentations. Each block contained 120 trials, and the order of attended versus distracted blocks was counterbalanced among participants. Assuming that the bug task consumed attentional resources to the extent that the scenes could not be fully processed, we expected to observe an amplitude reduction of the oddball-P3b ERP component for the dual-task, scene-distracted blocks relative to the single-task, scene-attended blocks (e.g., Sirevaag, Kramer, Coles, & Donchin, 1989). We note that this paradigm constitutes a conservative estimate of the degree to which the bug task draws resources away from the scenes, since here the participants are asked to perform a dual task. In the main experiment, participants are asked to ignore the scenes and concentrate fully on the bug task. We also looked for attentional enhancement of N1 and P2 sensory component amplitudes to scene presentations in the attended blocks, and for attentional enhancement of the P3b component to bug color changes in the distracted (bug game attended) blocks, across all blocks in the main experiment as evidence of attentional allocation differences between the attended and distracted conditions.

EEG data were collected from a Brain Products BrainAmp DC amplifier using 26 passive electrodes (Ag/AgCl) evenly spaced over the scalp (see insets in Figures 2–4), referenced to the left mastoid online and rereferenced to the arithmetic average of the left and right mastoids offline. Saccades and blinks were monitored by additional electrodes placed on the outer canthus of each eye and below the left eye. Impedances were lowered below 5 KΩ before the start of the experiment. Trials were time-locked to scene onsets. Data were sampled at 1000 Hz and online bandpass filtered between .016 and 250 Hz.

Analyses were performed using EEGlab and ERPlab. Data were bandpass filtered offline (Butterworth filter with 12 db/oct rolloff) between 0.1 and 50 Hz (half-amplitude cutoffs) and binned into 700-msec epochs including a 100-msec prestimulus baseline period. Epochs containing artifacts were rejected using ERPlab, where epochs containing blinks and saccades were automatically removed after being identified from derived channels capturing a polarity-inverted signal above and below the eye (characteristic of blinks) and on the left and right sides of the eye (for saccades), with thresholds set on a condition-blind but participant-tailored basis to account for variability in eye movement amplitudes and overall noise levels among participants. The modal blink threshold was 70 μV with a range of 40–100 μV, and the modal saccade threshold was 20 μV with a range of 15–40 μV, which resulted in a cumulative mean rejection rate of 15% of epochs per participant with a range from 1–39%.

We selected a temporal analysis window spanning from 250 to 400 msec to obtain per participant averages of N300 amplitudes, based on our pilot data. Following our preregistered procedure, before viewing individual condition averages, analysis electrodes were selected by creating a grand average of all conditions and selecting the electrode with the largest negative voltage deflection in our analysis time window from frontal, central, and parietal electrodes (because of the known fronto-central distribution of the N300 component; Schendan, 2019; Schendan & Kutas, 2002, 2003), then testing for significant differences in voltage between this maximally negative electrode and adjacent electrodes until a region boundary was defined by statistically significant decreases in voltage (paired-samples t tests with α = .05 and without correction for multiple comparisons). The electrodes that met this identification criteria were the right and left medial frontal electrodes (RMFr and LMFR; see insets of Figures 3A, 4A, or 4B), which were thus combined into an ROI that was used in all condition-based analyses, performed based on mean amplitudes measured from condition-averaged data across the 250- to 400-msec time window. Data collected during manipulation check blocks was processed in the same manner as described above with the exception that pooled data from the middle occipital, and right and left medial occipital electrodes (MiOc, RMOc, and LMOc; see inset of Figures 2A or 2B) were analyzed in a 250-to 500-msec time window to characterize the P3b (see inset of Figure 2). To test for attention-based effects on sensory components, we measured mean amplitudes at channels MiOc, RMOc, and LMOc from 100 to 150 msec for the N1 component and 175 to 225 msec for the P2 component. Analysis of a frontal P2 component was later added as a post hoc measure, which measured activity between 175 and 225 msec in pooled RMFr and LMFr channels.

We had three preregistered hypotheses with respect to the real-world statistical regularity manipulations, all of which were assessed via t tests:

1. N300 good/bad effect when attended: We predicted greater negativity for bad exemplars than good exemplars in the real-world statistical regularity attended condition (one-tailed), a replication of Kumar et al. (2021).

2. N300 good/bad effect when distracted: We predicted greater negativity for bad exemplars than good exemplars in the real-world statistical regularity distracted condition (one-tailed), indicating that full attention is not necessary for sensitivity to real-world statistical regularity.

3. Attention's impact on good/bad N300 effect: We predicted no statistical difference between attended and distracted conditions (two-tailed) for the real-world trials; if no difference was observed here, we would perform an equivalence test between the two conditions comparing the observed effect to a smallest effect of interest of dz = 0.3 (this exact effect size was chosen only as a general proxy for a small effect).

We additionally performed two tests for our attention manipulation checks with the following hypotheses, again assessed via t tests:

1. Attention impact on oddball P3 (manipulation check): We predicted greater positivity for oddballs than standards in the oddball task (one-tailed) and, critically regarding the manipulation check, a larger oddball P3b for scene-attended than scene-distracted conditions (one-tailed).

2. Attention impact on sensory components (manipulation check): We predicted a larger (more negative) N1 and larger (more positive) P2 amplitude for attended than distracted trials in main experimental blocks (one-tailed).

Finally, we performed the following exploratory tests with respect to artificial statistical regularity:

1. Joint impact of context and attention on N300 and N400: We performed a repeated-measures ANOVA with two levels of Representativeness (high and low), two levels of Attention (attended and distracted), and two levels of Context (real-world and artificial). Any significant interactions would be followed with t tests on the components of the interaction.

2. Impact of exemplar frequency: We predicted greater negativity for infrequent (20%) than frequent (80%) exemplar types, assessed via t tests. Such a result would indicate that artificial statistical regularity (built up by the experimental context) affects N300 amplitude.

Any nonpreregistered, post hoc t tests were computed as two-tailed tests. Standardized effect sizes accompany each set of inferential statistics. Our preregistered analysis planned to compare effect sizes to interpret the practical significance of our effects. Since the time of writing the preregistration, Bayes factors have become an increasingly popular tool for this purpose, particularly in assessing evidence for the null hypothesis, and we have accordingly added Bayes factors (denoted as BF) to each set of inferential statistics to provide additional context to our inferences. A standard Cauchy prior parameter of 1 was used in calculating Bayes factors as suggested by Rouder, Speckman, Sun, Morey, and Iverson (2009), and one-sided calculations against the point null were used where accompanying one-sided t tests were used (and, likewise, two-sided calculations against the point null were used where accompanying two-sided t tests were used). Bayes factors related to ANOVAs were calculated using the anovaBF function from the R package BayesFactor with the whichModel parameter set to bottom, which tests the effect of adding factors and factor interactions to the null model (Morey & Rouder, 2024). We adopt the convention of interpreting Bayes factors greater than 3 as evidence in favor of the alternative hypothesis and those less than 1/3 as evidence in favor of the null hypothesis, with values falling in between to be interpreted as non-informative or merely anecdotal.

Analyses for exclusion were conducted before any further analyses. Participants would be excluded from primary analyses if any of the criteria were met (selected values informed by distracted session pilot, n = 12, and separate attended session pilot, n = 3). Data collection continued until our desired sample size of 20 participants was acquired:

• Memory Quiz: Participants would be excluded if they scored less than 12 out of 18 correct on the memory quiz in scene-attended blocks (binomial test probability that exactly 12 or more correct responses of 18 corresponds to guessing, p = .12).

• Bug Task: Participants would be excluded if, on average, they were unable to keep the bug inside the bubble for less than or equal to 80% of the time in scene-distracted blocks.

• Eye Movements: Participants would be excluded if 40% or more trials needed to be excluded because of eye movements.

No participants met exclusion criteria on the basis of excessive eye movements or poor performance on the bug task. However, our memory performance criterion appeared to better index memory than attention, as it achieved the opposite of our desired effect by removing participants who displayed the largest attention effects for components well-established to be attention sensitive (see Appendix). Furthermore, because our pilot sample did not involve participant exclusion on the basis of the memory quiz, we might have overestimated effect sizes in comparison to our main sample, which did exclude participants on the basis of memory performance, causing a lack of statistical power to detect the predicted effects. Thus, we present results from the full set of 34 participants in the following sections, and the analysis limited to the smaller sample passing the memory quiz criterion is presented separately in the Appendix.

Average accuracy on memory quiz trials across all 34 participants was 67% (SD = 12%). Average performance on the bug task defined by percentage of frames the bug was contained within the bubble was > 99% (SD < 1%). When asked to rate the difficulty of the bug task on a scale from 1 to 10, the mean participant rating was 4.24 (SD = 1.77).

Manipulation check ERP results are plotted in Figure 2. The difference in P3b effects between single task (M = 3.24, SE = 0.62) and dual task (M = 3.86, SE = 0.58) in the manipulation check blocks was not significant, t(33) = −1.03, p = .84 (one-tailed), dz = −.18, BF = 0.37, indicating that the attentional resources consumed by the bug task were not sufficient to significantly disrupt the allocation of attention to the oddball task, although with only anecdotal support for this interpretation from the associated Bayes factor. This result does not ultimately provide support in favor of or against the success of the distraction manipulation, and we note that dividing attention between two tasks (bug and oddball) may not produce as large of an attention effect as an attention manipulation in which attention is directed toward one task (bug task) or the other (scene task), with instructions to ignore the non-task relevant information. Thus, instead of relying solely on the dual task oddball task (which took place after the main experiment blocks) to estimate the degree to which the bug task might have consumed attention, we also examine measures of attention that were collected while the main task occurred.

During the main task, the main effect of Attention (increased amplitude while attending) was significant for the posterior N1 component (attended: M = −0.16, SE = 0.50; distracted: M = 0.19, SE = 0.46), t(33) = 1.78, p = .043 (one-tailed), dz = .30, BF = 1.11, and the anterior (see Figure 6) P2 component (post hoc analysis; attended: M = −4.24, SE = 0.43; distracted: M = −5.53, SE = 0.38), t(33) = 6.59, p < .001 (two-tailed), dz = 1.13, BF = 97,205, but not for the posterior P2 component, t(33) = −1.72, p = .96 (one-tailed), dz = −.30, BF = 1.01. Although support from the Bayes factor for effects on the posterior N1 is rather weak, the evidence for the anterior P2 difference is very strong. Furthermore, across all main blocks, P3b amplitude was significantly greater in response to bug-attended than bug-ignored color changes (attended: M = 3.53, SE = 0.28; distracted: M = 0.12, SE = 0.16), t(33) = 11.57, p < .001 (one-tailed), dz = 1.98, BF = 5.30 × 10¹⁰.

Taken together, the findings support the conclusion that participants followed task directives: They attended to the bug game in the distracted condition but not the attended condition, as evidenced by the P3b difference to the bug color change, and they attended more to the scenes in the attended than the distracted task, as evidenced by the frontal P2. All Bayes factors either supported this interpretation or were indecisive, but none provided notable evidence in favor of a lack of attentional differences between the two tasks. We do not claim to have conclusive evidence that the bug task consumed resources to the extent that the scenes could not also be attended to some degree; however, we take the overall pattern of results to indicate that the distraction task prevented full attention from being allocated to the scenes, as intended.

Real-world statistical regularity ERP results are plotted in Figure 3. Bad exemplars (M = −6.56, SE = 0.39) elicited significantly larger N300s than good exemplars (M = −5.89, SE = 0.37) in the attended condition, t(33) = 4.71, p < .001 (one-tailed), dz = .81, BF = 1,040. Bad exemplars (M = −6.82, SE = 0.31) also elicited significantly larger N300s than good exemplars (M = −6.40, SE = 0.28) in the distracted condition, t(33) = 2.20, p = .018 (one-tailed), dz = .38, BF = 2.35, although its Bayes factor suggests only anecdotal strength. Moreover, the difference between attention conditions was not significant, t(33) = 1.10, p = .28 (two-tailed), dz = .19, BF = 0.24. The effect size of the difference between conditions was smaller than our smallest effect size of interest of dz = .3, although we note that the 95% confidence interval around our dz = .19 estimate is wide (−.15, .53) and a substantially larger sample would be required for a precise estimate. Nevertheless, it is worth noting that the Bayes factor here indicates moderate evidence for the lack of a difference between attended and distracted N300 effects. Thus, although we cannot rule out the possibility that attention has some small modulatory effect, we can infer that full attention is unlikely to be necessary for the brain to respond to representativeness in natural scene images.

The following sections describe the results of exploratory analyses that we feel are informative in contextualizing or extending our primary analyses, but which were not preregistered before the beginning of data collection (beyond exploratory Hypothesis 2 regarding the impact of exemplar frequency in the artificial statistical regularity blocks).

The order of conditions was split such that some participants experienced the attended condition first, whereas others began with the distracted condition. Is it possible, then, that those who began with the attended condition continued attending during the distraction condition and that this subgroup drove the overall pattern of results? As seen in Figure 3B, the distributions are not bimodal and the lack of an N300 effect difference between attended and distracted conditions does not appear to be driven by outliers. Nonetheless, we sought to explicitly rule out this possibility.

The number of participants beginning with either condition was equally balanced in the n = 20 sample who passed the memory quiz criterion, but on the way to arriving at that sample, more participants who began with the distracted condition happened to fail the memory quiz than those who began with the attended condition. Replacing those participants resulted in 14 participants who began with the attended condition and 20 participants who began with the distracted condition in the full n = 34 sample. This imbalance prevents testing order effects via a repeated-measures ANOVA, so instead, we will employ hierarchical linear modeling and use chi-square tests of log-likelihood to compare nested models to test whether condition order impacted our results.

First, we fit a null model to predict N300 mean amplitudes from the real-world statistical regularities condition with only a fixed intercept and a random intercept for participants as a level-two grouping variable. Adding a fixed effect for Representativeness (good/bad exemplars) significantly improved model fit, χ²(1, n = 34) = 10.14, p = .001. Keeping with other analyses, adding an interaction and main term for attention did not significantly improve model fit, χ²(2, n = 34) = 5.92, p = .052. We suspect that the reason adding Attention effects here comes close to a significant improvement stems from a combination of the excellent statistical power of hierarchical linear modelings and a carryover main effect of attention from the earlier P2 component. Inspecting Figure 3A, it appears that an effect of attention on the P2 component bled into the early portion (250–300 msec) of our preregistered analysis window, where the waveforms of attended conditions (solid lines) are less negative than those of the distracted conditions (dashed lines). However, this difference clearly disappears in the more N300-dominant section of the window (300–400 msec). We then added main and interaction terms for counterbalancing order, and again found no significant improvement in model fit, χ²(4, n = 34) = 0.99, p = .91, indicating that order effects are unlikely to play a role in the overall pattern of results.

A repeated-measures ANOVA revealed a significant main effect of Representativeness (larger N300s for bad than good exemplars), F(1, 33) = 101.7, p < .001, η_p² = .76, BF = 1.65 × 10¹⁰, and a significant main effect of Context (larger N300s in real-world statistical regularity conditions than artificial statistical regularity conditions), F(1, 33) = 20.51, p < .001, η_p² = .38, BF = 1,432, but again no significant main effect of Attention (attended/distracted), F(1, 33) = 1.01, p = .32, η_p² = .03, BF = 0.23. Also significant was the Representativeness × Context interaction, F(1, 33) = 9.56, p = .004, η_p² = .22, BF = 423.05, such that good exemplars elicited particularly small N300s in artificial statistical regularity conditions. The three-way interaction and remaining two-way interactions with Attention were nonsignificant (ps > .05). In summary, these results indicate larger N300s elicited by bad compared with good exemplars, larger N300s elicited in the real-world statistical regularity condition, and bigger good/bad effects in the artificial statistical regularity compared with the real-world statistical regularity condition, which were primarily driven by smaller responses to good exemplars in the artificial statistical regularity blocks. Yet, there was no significant effect of Attention on the N300.

Artificial statistical regularity ERP results are plotted in Figure 4. When good exemplars were more frequent (Figure 4A), bad exemplars elicited significantly larger N300s than good exemplars in both the attended condition (bad: M = −6.78, SE = 0.38; good: M = −4.86, SE = 0.36), t(33) = 9.24, p < .001 (one-tailed), dz = 1.58, BF = 213,680,353, and the distracted condition (bad: M = −6.51, SE = 0.31; good: M = −5.23, SE = 0.28), t(33) = 6.90, p < .001 (one-tailed), dz = 1.18, BF = 459,561. There was a significant difference between the attention conditions, t(33) = 2.34, p = .025 (two-tailed) dz = .40, BF = 1.60, such that the N300 effect was greater in the artificial statistical regularity condition when scenes were attended, although its Bayes factor indicates only anecdotal evidence in favor of an attention-related difference.

When bad exemplars were more frequent (Figure 4B), bad exemplars no longer elicited significantly larger N300s than good exemplars in the attended condition (bad: M = −5.89, SE = 0.38; good: M = −5.51, SE = 0.40), t(33) = 1.40, p = .085 (one-tailed), dz = .24, BF = 0.62, but the good/bad effect was reliable in the distracted condition (bad: M = −6.22, SE = 0.28; good: M = −5.61, SE = 0.34), t(33) = 3.20, p = .002 (one-tailed), dz = .55, BF = 20.80. However, the difference between attention conditions was not significant, t(33) = 0.69, p = .50 (two-tailed) dz = .12, BF = 0.17, with moderate evidence in favor of the null hypothesis as indicated by the Bayes factor.

Finally, when attending, the N300 good/bad difference was larger during contexts of frequent good exemplars than contexts of frequent bad exemplars, t(33) = 4.77, p < .001 (two-tailed), dz = .82, BF = 614, again suggesting that artificial statistical regularities can augment the effects of real-world statistical regularities on the N300. The N300 good/bad difference was also larger during contexts of frequent good exemplars than contexts of frequent bad exemplars when distracted, t(33) = 2.64, p = .013 (two-tailed), dz = .45, BF = 2.88, although the strength of this evidence was only anecdotal.

To explore the effects of our manipulations in the time window immediately after that used for the N300, encompassing more semantic aspects of processing on the N400 and late positive complex, we measured waveform amplitudes in a measurement window of 400–600 msec poststimulus using the same fronto-central ROI as in the analysis of the N300 (accounting for the more frontal distribution of the N400 for color images as opposed to words; West & Holcomb, 2002; Ganis, Kutas, & Sereno, 1996). We used a repeated-measures ANOVA with the same factors as those used in our analysis of the N300. The analysis revealed a main effect of Representativeness, F(1, 33) = 45.92, p < .001, η_p² = .58, BF = 1384, again such that bad exemplars elicited greater negativity than good exemplars, a main effect of Attention, F(1, 33) = 33.57, p < .001, η_p² = .50, BF = 4.69 × 10⁹, such that exemplars elicited greater negativity when attended than when distracted, and a main effect of Context, F(1, 33) = 93.58, p < .001, η_p² = .74, BF = 1.25 × 10¹⁰, such that exemplars from the real-world statistical regularities blocks elicited greater negativity than those from the artificial statistical regularities blocks. The Representativeness × Attention interaction, however, was not significant, F(1, 33) = 1.82, p = .19, η_p² = .05, BF = 0.13. Neither of the remaining two-way interactions or the three-way interaction were significant (ps > .05, BFs < 0.20). Thus, in this later time window, there was a general effect of attention that was not present for the N300, although still no evidence for an interaction between attention and representativeness.

In this experiment, we sought to investigate the role of attention in the brain's processing of real-world statistical regularity. Participants were exposed to a randomly shuffled series of unique natural scene images where each image was of high or low representativeness to its category, and were asked either to focus on the images or focus on playing a game instead while we recorded brain activity time-locked to the image presentations. We replicated previous work showing that the N300 ERP component is sensitive to manipulations of representativeness in natural scenes when they are attended, and demonstrated that patterns of N300 activity elicited by these scenes remain unchanged even when participants are distracted (Figure 3). In an exploratory analysis, we then found that adding additional artificial structure by changing the ratio of good and bad exemplars within a block further augmented these effects produced by the real-world statistical regularities present in the images, such that N300 effects were larger when blocks were primarily composed of good exemplars and interspersed with bad exemplars, and smaller when the reverse was true, in a manner that still did not vary convincingly with attention (Figure 4). Overall, the pattern of N300 amplitudes showed some sensitivity to artificial statistical regularities (experiment-dictated patterns altered the size of the good/bad effect), yet patterns remained consistent with respect to real-world statistical regularities across conditions; items of low representativeness continued to elicit larger N300 amplitudes, regardless of their relative frequency within a block (i.e., the effect did not flip such that good exemplars elicited larger N300s when in the minority; Figure 4B), supporting the N300 component's specificity to long-term as opposed to short-term expectations.

Although the N300 component was originally associated with gauging global image structure (Schendan & Kutas, 2002, 2003), recent theories have linked it to object model selection, in which the best match to the input is selected from memory (Schendan, 2019). Moving beyond object processing, we observe that a common thread across many different experiments showing N300 effects is that they all employ stimuli—including, as here, natural scenes—that are more or less statistically regular (Kumar et al., 2021; Schendan & Ganis, 2015; Gratton, Evans, & Federmeier, 2009; Schendan & Lucia, 2009; Schendan & Maher, 2009; Folstein, Van Petten, & Rose, 2008; Folstein & Van Petten, 2008; Gruber & Müller, 2006; Folstein & Van Petten, 2004; Schendan & Kutas, 2002, 2003; McPherson & Holcomb, 1999; Schendan, Ganis, & Kutas, 1998; Holcomb & McPherson, 1994); that is, the statistically regular stimuli all arguably contain information that will speed the processing of what the stimulus is. Thus, we have inferred from this body of previous evidence that the N300 indexes a match to prediction that is based on real-world statistical regularity. Schendan (2019) describes the N3 complex, of which the N300 is a dominant component, as involving implicit memory, attentional reorienting, automatic mental stimulation, and phenomenological consciousness, among other processes, and places the N3 complex in Stage 2 of her model, where higher order processes requiring feedback occur. Schendan does, however, acknowledge that even some processes thought to be “higher order” may happen automatically in some contexts. The most significant and novel contribution of this experiment is to show that focused attention is not critical in obtaining N300 effects, and thus neither, we argue, is it critical to the brain's processing of real-world statistical regularity, reinforcing the idea that even seemingly higher order processes like exemplar categorization or template matching may occur without attention.

The observed pattern of activity situates the N300 component in a unique position. Whereas effects on components like the P3b arise from situations that violate conscious, task-driven expectations (Polich, 2007) and disappear as attentional allocation falls below a threshold (Del Cul, Baillet, & Dehaene, 2007), and components like the vMMN arise from situations that violate subconscious stimulus-related expectations (Kimura, 2012), even (for the auditory MMN; Loewy, Campbell, & Bastien, 1996) in situations where participants are asleep, the N300 seems to occupy a middle ground, both temporally and functionally. Our data, in combination with previous investigations into the N300, have demonstrated that it displays greater amplitudes in situations that either consciously or subconsciously subvert our long-term expectations about incoming visual information. We think this activity reflects the automatic application of long-term knowledge in the service of processing incoming visual information, because the results obtain even when the upcoming stimulus cannot be predicted from the context of the experiment. Our results further illuminate the nature of the encoding of real-world statistical regularities in the brain and are most consistent with a view wherein these encodings are built into the early to middle stages of the visual processing hierarchy and act more like automatic continuous filters, rather than like comparators that are only brought online when explicitly needed. The processing of real-world statistical regularity, however, may still be biased by “online” predictions induced by recent patterns of sensory data (as in the artificial statistical regularity condition) and explicit top–down cues (as in the cued condition in Kumar et al., 2021).

Some have argued for a central role of attention in processing scenes, for example, suggesting that attention is necessary for forming a “virtual representation” of a scene (Rensink, 2000), or, conversely, that a “scene grammar” is automatically applied to a viewed scene, which then guides the focus of attention (Võ, Boettcher, & Draschkow, 2019). Either view is potentially consistent with our own data, as we see these “schema”-application processes as semantic in nature, whereas we see the processing of statistical regularities as mostly visual (and presemantic) in nature. Semantics might still influence how real-world statistical regularities are learned, but once learned, they are automatically evoked to help organize the visual input. Indeed, the N300 response reported here is sensitive to representativeness while it is relatively insensitive to attention. On the other hand, we found in an exploratory analysis that processing after the N300, reflecting semantic access and integration in the N400 and late positive complex components, responds both to representativeness and to attention (although without interacting; similar to the fMRI results obtained for PPA in Shao & Beck, 2024). We suggest that the representativeness of a scene helps organize the visual input so that it will more quickly make contact with semantics, after which attention provides a general boost to the semantic signal in later processing stages. Although Figure 3 suggests a trend such that the difference between good and bad exemplars seems to last longer when attended in the real-world statistical regularity condition, we did not find evidence for an overall interaction between representativeness and attention, casting doubt on the notion that these later semantic processes require attention to maintain distinctions in representativeness. Confirmatory experiments with targeted manipulations and more sensitive behavioral outcome measures will be needed to better address this relationship. At present, we conjecture that predictions based on real-world statistical regularities, associated with semantic concepts although not inherently semantic themselves, are employed without attention, and that attention may then boost that signal to be better leveraged in tasks that benefit from the use of semantic knowledge (i.e., scene categorization or visual search). This view is in line with those advocating a functional dissociation between the N300 and N400 wherein the components are thought to reflect separate processes of visual identification and semantic access, respectively (Truman & Mudrik, 2018; Hamm, Johnson, & Kirk, 2002; however, see also Draschkow, Heikel, Võ, Fiebach, & Sassenhagen, 2018).

The more automatic nature of these proposed prediction processes is also consistent with the literature on category information and scene gist being extracted without focal attention (Greene & Fei-Fei, 2014; Li et al., 2002). Indeed, the two processes, category/gist extraction and the predictive processes studied here, might be intimately related. For instance, the brain's propensity to predict may enable fast gist extraction. Alternatively, fast and early gist extraction may serve to initiate the prediction signal. In either case, the utility of these processes proceeding without focused attention is clear. The need to verify a percept's identity with focused attention would dramatically slow processes known to be exceedingly fast (i.e., gist extraction: Walther, Caddigan, Fei-Fei, & Beck, 2009; VanRullen & Thorpe, 2001; Thorpe et al., 1996; Potter, 1976). In addition, as described above, perceptual predictions and fast gist provide a framework to then more effectively employ attention to verify or act within the environment.

Although we did find that the N300 effect can be observed when attention is drawn away from the eliciting stimuli, toward another task, we cannot make the claim that stimuli in distracted blocks were wholly unattended. Our P2 and P3b main task results do support the case that participants allocated less attention to the stimuli in the distracted blocks than the attended blocks; however, these results can provide only a rather coarse window into the extent to which attention was withdrawn from the scenes. Moreover, although the ERPs elicited in the single versus dual task oddball blocks visually conform to the predicted pattern of largest amplitudes for single task oddballs, smaller amplitudes for dual task oddballs, and smaller amplitudes still for standards from both tasks within the middle 300- to 400-msec range of our P3b window, P3b amplitudes across the entire 200- to 500-msec window did not significantly differ between conditions, suggesting that the bug task potentially did not deplete attentional resources to the extent that we had hoped.

We should note that evidence from main task blocks and manipulation check blocks offer unique strengths and weaknesses. On the one hand, distraction away from presented images (the statistically insignificant effect on the bug task on the oddball P3b in the manipulation check) might more cleanly capture the process we wish to characterize than attention toward the distractor task (the statistically significant increase in P3b effects to attended over ignored bug color changes). On the other hand, the task in the manipulation check blocks was quite different than that of the main blocks; participants were explicitly asked to pay attention to repeated standard and oddball images that were trivial to distinguish, rather than to focus solely on playing the bug game while unique, random, task-irrelevant images flashed in the background, as in the main task. In this sense, the task in the manipulation check blocks might have been so easy that we arrived at a ceiling effect and could not sufficiently strain attentional resources to the point of seeing a difference in the P3b. In addition, because manipulation check blocks always occurred after all main blocks were completed, they are further removed from the brain processes we wish to qualify. It is even possible that the bug task consumed attentional resources to a great extent earlier on in the experiment but less so by the end of the experiment, when participants were well-practiced.

Future experiments may still consider exploring stronger distraction manipulations. Some experiments in scene gist perception have argued that tasks were not sufficiently difficult to draw significant attentional resources, which could have led to the acquisition of gist even in “unattended” conditions (e.g., Cohen et al., 2011). As some bit of assurance, Shao and Beck (2024) arrived at a similar pattern of results to those presented here and manipulated attention using a demanding RSVP task for which participants performed well but were not at ceiling (correct detection rate: 90.59 ± 1.89%; false positive rate: 8.62 ± 5.06%). Moreover, their RSVP task was previously shown to reduce activity elicited by unattended stimuli (Schwartz et al., 2005), further suggesting that resources were pulled away from unattended stimuli. In both studies, as a first pass, we only considered the question of whether attention is necessary to index real-world statistical regularity in an all-or-nothing sense, yet, as previously mentioned, there remains the possibility that attention could have a modulatory effect on the use of real-world statistical regularity. Subsequent designs could manipulate distractor task difficulty in a more granular manner to test this possibility. Although our overall pattern of results does not suggest a critical role of attention in eliciting N300 effects, we would not argue that attention plays no role in processing real-world statistical regularity. It remains possible that attention could have had subtle effects that were undetectable at our sample size. We do argue, however, that, in comparison to the magnitude of the overall effect of real-world and artificial statistical regularity on the N300, the contributions of attention seen here are at best small.

Although representations of real-world statistical regularity might be leveraged without focused attention, it is not the case that they are independent of top–down expectations. As noted, Kumar et al. (2021) found that when a specific category was explicitly cued into awareness using a text descriptor before the onset of the scene, the N300 for good and bad scenes differed only when the cue matched the upcoming scene and not when the cue indicated a different category than was ultimately presented. As argued above, this result is consistent with our interpretation of the N300 as reflecting a matching process with a perceptual prediction, which in this case is a prediction initially evoked by the cue rather than the features in the image. Under this interpretation, then, the cue impacts the expectation, or prediction, that the system brings online. Thus, this manipulation arguably reflects an effect of expectation, and, as mentioned above, a number of researchers have argued that expectation and attention reflect different mechanisms (Rungratsameetaweemana & Serences, 2019; Summerfield & Egner, 2009, 2016; Summerfield & De Lange, 2014; Feldman & Friston, 2010; Friston, 2009).

Unlike Kok et al. (2012) who describe a synergistic interaction between expectation and attention, we find little effect of attention and no interaction with expectation as indexed by the N300. Our experiments are very different in nature, as theirs sets up expectations within the context of the experiment and uses simple, artificial stimuli that will engage different brain processes, both because of their simplicity and the low probability that participants would have had any meaningful experience with them in the real world. A difference in measurement method could also contribute to the discrepancy, as Kok et al., (2012) used fMRI that relies on the slower BOLD signal that unfolds over the course of seconds rather than milliseconds, potentially averaging across multiple brain processes that may index statistical regularities, attention, and their interaction. We selected a temporally specific, functionally sensitive ERP measure in the N300 based on its link to real-world statistical regularity provided in the ERP literature, but it is possible that other sources of brain activity in earlier or later latency windows might show an expectation–attention interaction, and it is these signals that dominate across the longer mean window that is captured in the BOLD signal reported by Kok et al., (2012). Although not finding evidence of an interaction, Shao and Beck (2024) did find a main effect of attention in area PPA using an fMRI design similar to the one used here; an effect that was absent in our N300 signal.

Others using fMRI with more complex stimuli (e.g., isolated grayscale faces, Larsson & Smith, 2012; isolated objects, Richter & de Lange, 2019) have argued that expectation effects do not occur outside of attention, analyzing regions such as the face-selective fusiform face area and the object-selective lateral occipital cortex. Larsson and Smith (2012) used repeated presentations of stimuli within the context of an experiment to induce repetition suppression (Grill-Spector & Malach, 2001) and took this as their metric of expectation. It is plausible that repeating a stimulus and measuring the decrease in the neural response could capture fundamentally different mechanisms (Grill-Spector, Henson, & Martin, 2006) than measuring differences in neural responses to unique, novel stimuli (note as well that it is highly improbable that repetition suppression could explain our data, given that every scene is unique and drawn randomly from one of six different categories). Most importantly, though, none of the experiments reporting effects of attention on expectation examine the same type of expectation we are examine here, that is, effects of real-world statistical regularities (thought to specifically modulate N300 activity) that are built from meaningful, goal-oriented interactions with the environment. Rather, they manipulate arbitrary temporal associations that must be learned over the course of an experiment and only remain relevant within the context of that experiment; the objects and faces presented are not inherently any more expected by participants than any other potential objects and faces outside of the artificial confines of these experiments.

We argue that it makes good sense for the brain's architecture to undergo reorganization in response to stimuli that are encountered repeatedly throughout the lifetime and whose recognition provides real benefits to the organism, yet there is little impetus to enact sweeping structural changes in response to arbitrary associations learned over the course of an experiment, which are unlikely to be encountered again outside of the experiment. The temporal element of these paradigms must be emphasized further, as statistical regularity in the Richter and de Lange case is a temporal prediction, which is to say that in this case, the identity of the object on trial n − 1 predicts the identity on trial n. Therefore, if one is not attending to trial n − 1 and is unaware of its identity, then a prediction about trial n might not be generated. This type of prediction regarding temporal associations between arbitrary stimuli is fundamentally different than a prediction regarding inherent meaningful qualities of a stimulus. Larson and Smith admit that some “stubborn” prior expectations developed over longer durations might endure regardless of how attention is allocated (Yon, de Lange, & Press, 2019; Kok, Bains, van Mourik, Norris, & de Lange, 2016). We suspect that real-world statistical regularities fall into this category of prior expectations. Indeed, when using a design in fMRI that specifically depends on real-world statistical regularities, Shao and Beck (2024) do not find that perceptual predictions rely on the full allocation of attention.

To our knowledge, only the complementary study by Shao and Beck (2024) has assessed whether full attention is necessary for the brain's assessment of real-world statistical regularity in fMRI, and this study is the first to do so in a more time-resolved manner in EEG. Other studies have shown that statistical learning may occur without focal attention (Duncan & Theeuwes, 2020; Batterink & Paller, 2019; Turk-Browne et al., 2005), although this is a different question for many of the reasons described in the Introduction section. Specifically, this line of research asks whether the learning of spatial and temporal contingencies for relatively simple stimuli can take place outside of attention, whereas the present experiment and Shao and Beck (2024) ask whether complex visual regularities learned over the lifetime can be applied in the assessment of incoming stimuli without requiring focal attention. The artificial statistical regularities condition used here where stimuli were blocked by natural scene category could represent a kind of halfway point between assessments of real-world statistical regularities and those found in typical statistical learning paradigms, however. Although there are no transitional probabilities for a particular image, as is most often used in statistical learning paradigms, we do have a very high transitional probability (100%) that the image will be from the same category and an 80% chance it will be of similar representativeness. Thus, these blocks provide the chance for the brain to pick up on regularities formed within the context of the experiment and to thereby make predictions about the upcoming category and diagnostic features. Although we have a strong selective attention task that might make acquiring this information more difficult (Turk-Browne et al., 2005), the regularities are fundamentally visual in nature and carry strongly ingrained priors that were learned over long time periods based on meaningful interactions with the world, potentially making them more accessible even outside of attention (Duncan & Theeuwes, 2020; Batterink & Paller, 2019; Yon et al., 2019).

Effect patterns in the artificial statistical regularity conditions suggest that the N300 is sensitive not only to long-term encodings of real-world statistical regularity, but also that these encodings can be molded by short-term constraints (Smith & Federmeier, 2020, 2024), even without full attention being allocated to those constraints, similar to findings from the statistical learning experiments cited above. Indeed, our largest N300 effects came from the artificial statistical regularity conditions where more good exemplars were presented than bad ones. We postulate that differences in real-world statistical regularity are magnified when the brain has a more highly constrained prediction/template space in which to evaluate exemplars. Good exemplars within a category are visually more similar to one another than bad exemplars (Caddigan et al., 2017; Torralbo et al., 2013), and this rich bed of consistency seems to provide an anchor that results in greater sensitivity to violations of perceptual predictions. Bad exemplars continued to elicit larger N300s even when they served as the majority class within a block, indicating a bias for the assessment of representativeness toward the long term over the short term. We note, however, that as our artificial statistical regularity results were exploratory, they should be confirmed in future experiments before firm conclusions are drawn.

We conclude that full attention is not critical in the brain's assessment of real-world statistical regularity. Here, we support this claim on the basis of our manipulation of one type of real-world statistical regularity, the representativeness of natural scenes, but have made the case that many previous ERP experiments using their own unique manipulations have shown the same variation in brain responses under full attention, all of which fall under the umbrella of manipulations to real-world statistical regularity. We therefore predict that these other paradigms would show similar robustness to distraction. Given the full body of evidence, we interpret attention as having a minimal role in processing real-world statistical regularities, whereas predictions acquired throughout the lifetime that serve useful roles in making sense of what is out there in the world play a larger role. Our findings suggest that selection of relevant real-world statistical regularities is a process that happens automatically and ubiquitously, where signals cueing deviations from these encodings do not require explicit top–down influence, but instead flow readily from the experience-tuned architecture of the visual hierarchy, squaring nicely with the types of perceptual predictions posited by predictive coding accounts of brain function.

The following sections describe results from the initial sample of 20 participants who survived the memory quiz inclusion criterion and provide strong motivation for discarding the criterion and instead analyzing results from the full sample. In the smaller sample, within real-world statistical regularity blocks, we replicated the N300 effect to good and bad exemplars observed previously under attended conditions. Critically, as hypothesized, the difference between the attended and distracted good/bad exemplar effects was not statistically significant. However, when tested directly within the distracted condition alone, we did not find a significant N300 difference between good and bad exemplars. Within artificial statistical regularity blocks, we observed an N300 difference under both attended and distracted conditions, and no significant difference between these effects. A closer examination ultimately prompted us to investigate data from the full set of 34 participants rather than relying on this smaller sample.

In total, we tested 34 participants to obtain 20 who passed exclusion criteria. All 14 exclusions occurred on the basis of memory quiz performance, whereas there were no participants excluded because of bug task performance or high rates of eye movements. We note that this memory quiz criterion was much more conservative than we had intended (as it was one facet of the experiment we had not pilot tested). The results obtained from the remaining 20 participants are described below.

Our criterion for memory quiz performance was 12/18 trials correct, or 67%. Average accuracy on memory quiz trials for the included participants was 75% (SD = 6%). Average performance on the bug task defined by percentage of frames the bug was contained within the bubble was >99% (SD < 1%). When asked to rate the difficulty of the bug task on a scale from 1 to 10, with 1 being easy and 10 being difficult, the mean participant rating was 4.53 (SD = 1.81).

We included an oddball dual task to assess the degree to which the bug task drew attention away from the scenes. We note first, however, that within the main experiment, we have clear evidence that participants were doing something different in the distracted condition than in the attended condition, in the form of a large anterior P2 difference in response to scenes on the basis of attention in real-world statistical regularity blocks (Figure 6), where P2 responses to attended exemplars (M = −4.28, SE = 0.56) were more positive than those to unattended exemplars (M = −6.07, SE = 0.52), t(19) = 6.81, p < .001 (two-tailed), dz = 1.52, BF = 12,909. Attended exemplars (M = −4.56, SE = 0.63) also exhibited larger P2s than unattended exemplars (M = −5.46, SE = 0.53) in artificial statistical regularity blocks (Figure 7), t(19) = 2.56, p = .019 (two-tailed), dz = .57, BF = 2.55, although the associated Bayes factor indicates merely anecdotal evidence.

The comparison of ERPs elicited by the oddball task under single and dual task conditions (i.e., while performing the bug task), plotted in Figure 5A, can give some indication of the degree to which the bug task draws resources away from the scene task. At least when participants are asked to share resources between tasks in the final set of manipulation check blocks, it would appear that the bug task drew negligible resources away from the oddball task. The difference in magnitude of the P3b effect between the single (M = 2.59, SE = 0.81) and dual task (M = 3.19, SE = 0.66) manipulation check blocks was not significant, t(19) = −.70, p = .75 (one-tailed), dz = −.16, BF = .32. Similarly, the main effect of attention on the posterior N1 component in main blocks (attended: M = −0.18, SE = 0.71; distracted: M = −0.13, SE = 0.58) was not significant, t(19) = .18, p = .43 (one-tailed), dz = .04, BF = .20, nor was the main effect of attention on the posterior P2 component in main blocks (attended: M = 3.47, SE = 0.70; distracted: M = 3.60, SE = 0.77), t(19) = −.42, p = .66 (one-tailed), dz = −.09, BF = .25, when we restrict the analysis to the preregistered 20 participants. Bayes factors indicate weak to moderate evidence for the null hypotheses in these comparisons.

However, as noted above, the anterior P2 effect was both statistically strong and visually obvious in the data from the main experiment. We argue that although this is a post hoc measure, it could serve as better indicator of attentional resource allocation than the P3b from manipulation check blocks given that it was acquired during the same blocks as measures of the N300, as opposed to blocks that took place at the end of the experiment where participants were asked to perform a different task.

Real-world statistical regularity ERP results are plotted in Figure 6. As predicted, bad exemplars (M = −6.86, SE = 0.50) elicited significantly larger N300s than good exemplars (M = −6.23, SE = 0.49) in the attended condition, t(19) = 3.33, p = .002 (one-tailed), dz = .74, BF = 22.78. In contrast to the pattern in the pilot data, bad exemplars (M = −7.02, SE = 0.46) did not elicit significantly larger N300s than good exemplars (M = −6.64, SE = 0.38) in the distracted condition, t(19) = 1.39, p = .090 (one-tailed), dz = .31, BF = 0.75. Importantly, however, the difference between conditions was also not significant, t(19) = 0.89, p = .39 (two-tailed) dz = .20, BF = 0.25. Thus, we did not find reliable evidence that attention modulates the N300 effect, although the lack of a significant effect within the distracted condition itself weakens the conclusion that attention does not impact the good/bad N300 difference. The effect size of the difference between conditions was smaller than our smallest effect size of interest of dz = .3. Although, we should also note that the 95% confidence interval around our dz = .20 estimate is wide (−.25, .64) and a substantially larger sample would be required for a precise estimate. Bayes factors were indecisive regarding the lack of an N300 difference in the distracted condition and provided moderate evidence in favor of the null hypothesis that the attended and distracted conditions did not truly differ in the magnitude of their N300 effects. Thus, overall, the data pattern seems to weigh against the claim that attention impacts the good/bad N300 difference.

Here, we move beyond our preregistered hypotheses and look to analyze the impacts of all factors present in the data. A repeated-measures ANOVA revealed a significant main effect of Representativeness (good/bad), F(1, 19) = 62.68, p < .001, η_p² = .77, BF = 150,587, with bad exemplars eliciting larger N300s than good exemplars. There was also a significant main effect of Context (real-world/artificial), F(1, 19) = 10.02, p = .005, η_p² = .35, BF = 15.84, as N300 responses were overall larger in real-world statistical regularity blocks. However, there was not a significant main effect of Attention (attended/distracted), F(1, 19) = .22, p = .64, η_p² = .01, BF = 0.10, again indicating a lack of a role of attention in contributing to N300 amplitudes, with the associated Bayes factor strongly supporting the null hypothesis. The Representativeness × Context interaction was significant, F(1, 19) = 7.18, p = .015, η_p² = .27, BF = 25.22, such that good exemplars elicited particularly smaller N300s under artificial statistical regularity conditions. The three-way interaction and remaining two-way interactions with attention were all nonsignificant (ps > .05).

Artificial statistical regularity ERP results are plotted in Figure 7. Bad exemplars (M = −7.04, SE = 0.52) elicited significantly larger N300s than good exemplars (M = −5.20, SE = 0.49) in the attended condition when good exemplars were more frequent, t(19) = 7.62, p < .001 (one-tailed), dz = 1.70, BF = 114,068, which is not surprising given our earlier results both here (in the real-world condition) and in previous experiments (Kumar et al., 2021). We would not expect more frequent good exemplars to diminish their advantage. Interestingly, bad exemplars (M = −6.78, SE = 0.41) elicited significantly larger N300s than good exemplars (M = −5.57, SE = 0.38) in the distracted condition also, t(19) = 4.73, p < .001 (one-tailed), dz = 1.06, BF = 417. This result again suggests that, despite the nonsignificant good/bad difference for the distracted real-world condition, the effects of attention on the sensitivity to real-world statistical regularity are minimal. In keeping with this conclusion, the difference between attention conditions was not significant, t(19) = 1.40, p = .10 (two-tailed) dz = .38, BF = 0.64, although the strength of evidence for the lack of a difference here is only anecdotal.

The blocks in which the bad exemplars were more frequent are of particular interest, as in these blocks artificial statistical regularity works against the patterns established by real-world statistical regularity. If artificial statistical regularity overrides or is stronger than real-world statistical regularity, then we would expect the good/bad N300 difference to disappear or diminish. The good/bad difference did not disappear; bad exemplars elicited significantly larger N300s than good exemplars in both the attended condition (bad: M = −6.35, SE = 0.54; good: M = −5.58, SE = 0.61), t(19) = 2.00, p = .030 (one-tailed), dz = .45, BF = 1.89, and the distracted condition (bad: M = −6.42, SE = 0.39; good: M = −5.77, SE = 0.46), t(19) = 2.83, p = .005 (one-tailed), dz = .63, BF = 8.48, suggesting that real-world statistical regularity effects maintain even when artificial structure is introduced. Bayes factors indicated anecdotal strength regarding for a good/bad difference in the attended condition whereas evidence for a difference in the distracted condition was moderate. As in the blocks with frequent good exemplars, the difference in the good/bad exemplar effect between attention conditions was again not significant here in the blocks with frequent bad exemplars, t(19) = 0.24, p = .81 (two-tailed) dz = .05, BF = 0.18, this time with moderate evidence for a true lack of a difference between effects in these conditions.

Finally, when attending, the N300 good/bad difference was larger within contexts with frequent good exemplars than within contexts with frequent bad exemplars, t(19) = 3.27, p = .004 (two-tailed), dz = .73, BF = 10.21, suggesting that both artificial and real-world statistical regularity can impact the N300. However, the same was not true when distracted: Contexts with frequent good exemplars did not produce larger good/bad differences than contexts with frequent bad exemplars, t(19) = 1.71, p = .10 (two-tailed), dz = .38, BF = 0.64, although the Bayes factor reveals that these data are ultimately indecisive.

With regard to our confirmatory analyses on real-world statistical regularity conditions, with this sample of participants, we (1) replicated our pilot results showing that bad exemplars elicit larger N300s than good exemplars when participants are attending, (2) failed to replicate our pilot results showing that bad exemplars elicit significantly larger N300s than good exemplars when participants are distracted, but (3) also found no significant difference between these effects, suggesting that it is unlikely that the attentional manipulation had a strong impact on the N300 patterns. The analyses of the artificial conditions further support the idea that attention does not have a notable impact on the N300 good/bad difference. There was an N300 good/bad difference in both attended and distracted conditions within artificial statistical regularity conditions, with no significant effect of attention on this difference, corroborated by our 2 × 2 × 2 within-participant ANOVA, which suggests significant main effects of representativeness (larger N300s for bad exemplars) and Context (larger N300s for real-world statistical regularity blocks), but no significant interactions or main effects related to Attention.

Most of the conclusions from this sample thus suggest that N300 effects can be observed without focused attention to the stimuli and that, indeed, attention has little impact on the size of the good/bad effect. Thus, one possibility is that the failure to detect a reliable N300 difference in the real-world distracted condition arose because of a lack of power. We excluded many more participants than we had anticipated as participants struggled to succeed on the memory quizzes. We did not base our memory task exclusion criterion on pilot data, and it would seem that the task was more difficult than we intended and may not have achieved the desired result of sorting those who were on task from those who were not. The fact that lure scenes were derived from practice blocks, giving participants prior exposure to them, and that the target scene was one of 60 scenes presented in the immediately previous block, likely made the memory quiz unduly challenging. In retrospect, it seems highly likely that the memory quiz performance better loaded onto memory than attention, and thus that the excluded participants did allocate attention in the desired manner, but simply failed to encode the stimuli well enough to discriminate them from similar stimuli later.

Fortunately, there are other, likely more sensitive, indices of attentional allocation than the memory quizzes that are present in the data. In particular, in addition to the task-concurrent anterior P2 differences on the basis of attention described above, there were notable P3b differences in response to the color change in the bug task across attended and distracted conditions (Figure 5B). Recall that the bug task was present in both attended and distracted conditions, but only fell under player control in the distracted condition. This design allows us to compare P3b responses to bug color changes in both conditions to get a sense of how well participants allocated attention to the task at hand within each condition. How do included and excluded participants compare on this metric? Contrary to the conclusions we might draw from the dual-task data, both groups displayed a larger P3b in response to attended bug color changes (included group, attended: M = 3.71, SE = 0.31, distracted: M = 0.18, SE = 0.24, t(19) = 9.52, p < .001 (one-tailed), dz = 2.13, BF = 2,770,650; excluded group, attended: M = 3.27, SE = 0.52, distracted M = 0.04, SE = 0.18, t(13) = 6.55, p < .001 (one-tailed), dz = 1.75, BF = 3,044). The difference between groups in P3b effect magnitude was not significant, t(26.15) = .49, p = .63 (two-tailed), dz = .17, BF = 0.28. Moreover, when we turn to our other manipulation checks derived from the main task, we find that there was even a significant main effect of Attention for the excluded group on the posterior N1 component (attended: M = −0.13, SE = 0.69; distracted: M = −0.66, SE = 0.75), t(13) = 3.26, p = .003 (one-tailed), dz = .87, BF = 15.98, although not for the posterior P2 component (attended: M = 3.96, SE = 0.72; distracted: M = 4.71, SE = 0.78), t(13) = −2.41, p = .98 (one-tailed), dz = −.64, BF = 3.88. Furthermore, we also found in a post hoc analysis a significant main effect of Attention on the anterior P2 component in the excluded group (attended: M = −3.92, SE = 0.63; distracted: M = −5.33, SE = 0.57), t(13) = 6.63, p < .001 (two-tailed), dz = 1.77, BF = 1,700.

If our memory quiz criterion was truly sensitive to attention, we would have expected to see significant attention boosts for these manipulation checks only for the included group; however, we observe large attention effects for both groups, and the largest effects in the excluded group. In short, the data provide evidence that both groups followed the instructions to direct attention toward or away from the bug task, and, more interestingly, in the main task, the effect of directing attention toward the bug task had a more noticeable effect on scene processing for the excluded group than the included group. Given the full body of evidence, we argue that the use of the full sample of 34 participants is justified and offers the best case for evaluating our original hypotheses.

Corresponding author: Evan G. Center, Department of Information Technology and Electrical Engineering, University of Oulu, Erkki Koiso-Kanttilan katu 3, room TS 369, Oulu, Northern Ostrobothnia 90570, Finland, or via e-mail: [email protected].

Experiment scripts, participant-level mean ERP component amplitudes, and analysis scripts are available at https://osf.io/kwszv.

Evan G. Center: Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Software; Visualization; Writing—Original draft; Writing—Review & editing. Kara D. Federmeier: Conceptualization; Funding acquisition; Methodology; Resources; Supervision; Writing—Review & editing. Diane M. Beck: Conceptualization; Funding acquisition; Methodology; Resources; Supervision; Writing—Review & editing.

This work was supported by a 2018 Beckman Institute Graduate Fellowship awarded to E. G. C.; the Office of Naval Research Multidisciplinary University Research Initiative (https://dx.doi.org/10.13039/100000006), grant number: N000141410671 to D.M.B.; the National Institutes of Health (https://dx.doi.org/10.13039/100000002), grant number: R01 AG026308 to K. D. F.; the European Research Council (https://dx.doi.org/10.13039/501100000781), grant number: ERC AdG, ILLUSIVE: Foundations of Perception Engineering, 101020977 to Steven M. LaValle; and Business Finland (https://dx.doi.org/10.13039/501100014438), grant number: HUMOR 3656/31/2019.

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance. The authors of this paper report its proportions of citations by gender category to be: M/M = .443; W/M = .203; M/W = .165; W/W = .190.