|Home | About | Journals | Submit | Contact Us | Français|
Previous work on visual search has suggested that only a single attentional template can be prioritized at any given point in time. Grouping features into objects and objects into categories can facilitate search performance by maximizing the amount of information carried by an attentional template. From infancy to adulthood, earlier studies on perceptual similarity show that consistent features increase the likelihood of grouping features into objects (e.g., Quinn & Bhatt, 2009) and objects into categories (e.g., shape bias; Landau et al., 1988). Here we ask whether lower-level intra-item similarity facilitates higher-level categorization despite inter-item dissimilarity. Adults participated in four visual search tasks where targets were defined as either one item (a specific alien) or a category (any alien) with similar features (e.g., circle belly shape, circle back spikes) or dissimilar features (e.g., circle belly shape, triangle back spikes). Using behavioral and neural measures (i.e., N2pc ERP component which typically emerges 200 milliseconds post-stimulus), we found that intra-item feature similarity facilitated categorization, despite dissimilar features across category items. Our results demonstrate that feature similarity builds novel categories and activates a task-appropriate abstract categorical search template. In other words, grouping at the lower item level facilitates grouping at the higher category level, which allows us to overcome efficiency limitations in visual search.
Previous visual search studies have shown that search for one item is more efficient than search for two or more items, as reflected in behavioral (reaction time and accuracy) and EEG measures (N2pc ERP component) (see Olivers et al., 2011; Grubert & Eimer, 2013; Nako, Wu, & Eimer, 2014; Nako, Wu, Smith, & Eimer, 2014). The N2pc is the established ERP marker of attentional target selection and emerges approximately 200 milliseconds after stimulus onset at electrodes contralateral to the hemifield of the target (e.g., Eimer, 1996; Luck & Hillyard, 1994). Using the N2pc to measure early search efficiency, Nako et al. (2014a) presented participants with 4-item search arrays containing different letters, one of which sometimes was the target. Participants were asked to search for one letter target (e.g., the letter A), two letter targets, or three letter targets among other letter distractors. The N2pc amplitude became attenuated (as did behavioral measures) as the number of potential targets increased. More efficient search for one vs. two or more items supports an account that limits attentional prioritization to only one target template at a time (Olivers et al., 2011). Attentional templates are working memory representations based on features guiding visual search that are deployed early in the information processing stream.
However, multiple target search can become more efficient if the set of target items form a natural or learned category. Nako et al. (2014a) also showed that search for any letter (i.e., a category target) among number distractors yielded N2pc components that were very similar to 1-item search (see also Egeth, Jonides, & Wall, 1972 for similar behavioral evidence). In other words, once participants were able to use category knowledge, searching for many items (e.g., 26 letters) became similar to searching for one item, whereas multi-item search in which a category-level template could not be deployed was much less efficient. This finding suggests that categorization, and grouping in general, can overcome the attentional limitation of efficient search for only one item at a time.
In summary, while attentional templates can contain a single low-level target-defining feature (e.g., one orientation, shape, or color) or one whole object (see Olivers et al., 2011), findings by Nako et al. (2014a) show that attentional templates also can consist of higher-level abstract features, such as categories composed of multiple objects (see additional evidence in Wu et al., 2013; Wu et al., 2015). Thus, grouping multiple features into one object and multiple objects into one category may be an effective strategy for maximizing the information content of one attentional template and overcoming attentional limitations during search. Indeed, an increasing number of studies show that searching for multiple features belonging to one object or multiple objects belonging to one category is efficient (for a review, see Cunningham & Wolfe, 2014). Thus, efficient grouping strategies (similar to ‘chunking’ strategies, e.g., Gobet et al., 2001) enable adults to maximize attentional search capacities (Treisman, 1982). While it is clear that objects in known categories (e.g., letters, clothing, human faces) can be grouped together, what strategies can unify objects from novel categories to facilitate visual search?
Previous work on the effects of perceptual similarity from infancy to adulthood shows that identical features increase the likelihood of grouping elements into larger units (e.g., X’s vs. O’s, Quinn & Bhatt, 2009) and objects into categories (e.g., shape bias; Landau et al., 1988). Moreover, perceptual similarity among targets has been shown to facilitate visual search among distractors, and even automatically capture attention in task-irrelevant ways (e.g., Duncan & Humphreys, 1989). If items composed of similar features are detected more efficiently, then a category of items containing matching features should produce more efficient search, even if the matching features are dissimilar across objects (i.e., even if inter-item dissimilarity is high). However, it remains unclear whether low-level intra-item similarity facilitates higher-level categorization, and in turn boosts the efficiency of category-level visual search. While perceptual similarity within objects may facilitate search for specific objects, perceptual dissimilarity between objects may hinder categorization.
Although counter-intuitive from the broader perceptual similarity literature, previous behavioral studies with infants and adults on feature correlation predict that intra-item feature similarity should facilitate categorization (e.g., Austerweil & Griffiths, 2011; 2013). Feature correlation studies have shown that objects with correlated features (e.g., bananas tend to be yellow and have a crescent shape) rather than uncorrelated features (e.g., jelly beans come in many colors) enable more robust representations of multi-part objects and categories. Having stronger individual object representations due to consistent feature correlations aids categorization of these objects, because the features determining category diagnosticity are more reliable and thus less corrupted by noise (Austerweil & Griffiths, 2011; 2013; Goldstone, 2000; Younger & Cohen, 1986). Consistent feature correlations lead to narrower generalization based on specific features of existing category members, suggesting a tighter category boundary compared to that of uncorrelated features. Unfortunately, establishing new feature correlations in a visual search task requires extensive training, which in turn constrains the number of targets that can be used in a given experiment. Furthermore, creating similar enough target and distractor stimulus sets to avoid low-level pop-out effects typically precludes true feature correlations (e.g., features on the target are shared by distractors, as in conjunction search). An alternative to creating new feature correlations is to ask whether feature similarity serves to facilitate category-based target search. That is, does intra-item similarity among a set of targets that define a category, despite inter-item dissimilarity, facilitate categorization by highlighting between-category differences?
There are two aims in the present study: 1) to determine whether a category of objects with perceptually similar features, despite dissimilar features across objects, leads to more efficient search for both individual items and for categories (compared to perceptually dissimilar features within an object), and 2) to determine whether this efficient search is due to the ability to group objects with similar intra-item features into a single categorical attentional template. While the first aim can be addressed easily with behavioral measures, the second aim can best be addressed with the N2pc ERP component. This component provides a window into the grouping mechanism underlying better behavioral performance because it is modulated differentially by the number of categorically relevant and irrelevant search targets. Specifically, the amplitude of the N2pc decreases in a non-linear manner as the number of related targets increases, and decreases linearly with unrelated targets (e.g., Nako et al., 2014a; Nako et al., 2014b). That is, searching for two or more unrelated items reduces the amplitude of the N2pc linearly compared to searching for one item. However, if these items are exemplars from the same category, the amplitude of the N2pc does not fall off linearly with the number of items in the category (e.g., search for any letter is similar to search for 2 specific letters, Nako et al., 2014a). Once features are grouped into objects and the objects are grouped into a category, the N2pc is present regardless of the number of items in the category. As a result, the non-linear fall off of the N2pc with increasing set-size can be used as diagnostic of a category-level attentional template. Importantly, the N2pc is modulated by grouping exemplars into a category, not by task difficulty (Wu et al., 2015).
Previous work has shown that the amplitude of the N2pc is modulated by other factors, such as number of targets (e.g., Mazza & Caramazza, 2012), distractors semantically related to the target (e.g., Telling et al., 2010), reward salience (e.g., Kiss et al., 2009), and low-level pop-out effects (e.g., Theeuwes, 2010, although this point is heavily debated with the other side arguing that top-down factors are the main contributors to the N2pc component, e.g., Eimer, 2014). The current study controlled for all of these previously identified factors in order to draw conclusions about grouping based on intra-item similarity or dissimilarity. Specifically, only one target (or non-target) was presented during a trial with a distractor, novel targets were used to avoid semantic processing, and the two categories did not differ on reward salience or bottom-up visual features (i.e., the same features were used across both categories). Although multiple factors are known to influence the amplitude of N2pc, here we controlled for these previously identified factors by comparing an item set (“category”) whose exemplars shared high intra-item similarity, with a set whose exemplars did not. Intra-item similarity was the only factor that differed between the two item sets. Our hypothesis based on our previous work is that after controlling for previously identified factors that contribute to the presence of the N2pc, if there is an N2pc during search for multiple targets, the items can be grouped into one set that is “categorical” in the sense that it is more abstract than the perceptual characteristics of the individual exemplars in the set. This hypothesis is primarily based on our previous work showing the presence of the N2pc during category search for a wide range of familiar category (e.g., letters, numbers, kitchen items, clothing, faces).
The present study measured behavioral and neural (N2pc) responses in four visual search tasks (2×2 design) where targets were defined as either one item (a specific alien) or a category (any alien) with similar intra-item features (e.g., circle belly shape, circle back spikes, Family S, Figure 1) or dissimilar intra-item features (e.g., circle belly shape, triangle back spikes, Family D). Critically, the two categories differed only in feature similarity within one alien, not in the set of actual features presented across categories. Based on the feature correlation literature, we predicted that intra-item similarity would facilitate categorization, despite inter-item dissimilarity, producing a large N2pc for Family S but no N2pc for Family D during Category search. Based on the perceptual similarity literature, we predicted that similar features could be chunked into one attentional template during Exemplar search, and therefore would produce larger N2pc amplitudes compared to items in Family D with dissimilar features. Finally, based on previous N2pc studies (e.g., Nako et al., 2014b; Wu et al., 2013), it is entirely expected that the N2pc amplitude would be larger for Exemplar (specific item) search compared to Category search, due to more precise template matching during Exemplar search. However, the critical comparisons were between (rather than within) the two categories differing only in intra-item similarity.
Sixteen adults (M = 23.75 years, SD = 4.81, range: 19-33 years, 10 females) participated in this study. Data from an additional 5 participants were excluded from the final analyses due to excessive eye movements (> 50% trials excluded). This is the typical inclusion rate and number of participants in N2pc EEG studies (e.g., Nako et al., 2014a). All participants were compensated $25 at the end of the study.
A novel stimulus set of cartoon alien figures, Wusters (www.callalab.com, Figure 1), was employed to control for perceptual differences between categories and to reduce the amount of previous knowledge employed in the task. Two categories of aliens with identical bodies differed only in the shape comprising the spikes on the back of the alien as well as the shape displayed on the belly. Eight shapes were used, including a circle, triangle, square, pentagon, hexagon, star, heart, and an X. In one category (Family S, Figure 1, top panel left), each alien had one of the eight shapes both on its belly and as its spikes. For example, if a circle appeared on the belly, that alien’s spikes would also be circles. The dissimilar family (Family D) consisted of a random set of eight belly shape-spike combinations: aliens with a square belly shape and triangle spikes, heart belly with hexagon spikes, pentagon belly with circle spikes, triangle belly with heart spikes, star belly with X spikes, X belly with square spikes, hexagon belly with star spikes, and circle belly with pentagon spikes. The images subtended 5.15°× 2.86°.
Subjective ratings of similarity were obtained from a separate group of 12 participants (M = 20.92 years, SD = 1.93, range: 18-26 years, 10 females) to confirm the intra-item similarity and inter-item dissimilarity for Family S and the intra-item and inter-item dissimilarity for Family D. For the individual aliens, participants were asked, “How similar is the shape on the alien's belly to the shapes on the alien's back?”, and for the pairs of aliens, participants were asked, “How similar are these two aliens?”. Participants had to report their similarity judgments on a scale from 1 to 5, 1 being not similar at all and 5 being extremely similar. Figure 2 shows the mean ratings for individual aliens (8 aliens per category), as well as for pairs of aliens. These ratings confirm our expectations: 1) For Family S, intra-item similarity was high, while inter-item similarity was low (Figure 2, first two bars), and 2) for Family D, both intra-item and inter-item similarity were low (Figure 2, last two bars). These results indicate that Family S contained aliens with high intra-item similarity and low inter-item similarity, while Family D contained aliens with both low intra- and inter-item similarity.
In this within-subjects design, each participant completed four tasks: 1) search for a specific alien in Family S (Exemplar search), 2) search for a specific alien in Family D (Exemplar search), 3) search for any alien in Family S (Category search), and 4) search for any alien in Family D (Category search). Participants completed 28 blocks of trials in total, with seven continuous blocks of trials for each of the four tasks. Task order was counterbalanced across all participants with two Latin squares so that half of the participants received two exemplar searches in a row followed by two category searches or vice versa, while the other participants received alternating exemplar and category search blocks. To minimize interference effects per participant, the exemplar targets for both search tasks were pseudo-randomized with the constraint that the exemplars from Families S and D did not have overlapping shapes between them. For example, if a participant searched for an alien with a circle belly shape and circle spikes (Exemplar from Family S), she also searched for an alien with a triangle belly shape and heart spikes from Family D, rather than the alien with the circle belly shape and pentagon spikes. This constraint was implemented to reduce confusion and task difficulty for an already difficult task. For both Exemplar search tasks, the same target was displayed at the beginning of every block. Participants were shown the complete inventory of 16 aliens split by category at the beginning of the experiment.
There were four trial types across the four search tasks (Exemplar S, Exemplar D, Category S, and Category D): Exemplar Match, Category Match, Foil, and No Target trials (Figure 1, bottom panel). In each Exemplar Search task, each of the seven blocks presented 28 Exemplar Match trials (specific target alien appeared in the search array), 28 Foil trials (non-target alien from the same category as the target alien), and 6 No Target trials (only aliens from the other family appeared). In the Category Search task, there were 28 Category Match trials (any alien from the target family appeared) and 28 No Target trials (only aliens from the other family appeared) in each block. Foil trials were included in the Exemplar search task to ensure that participants searched for an exact item, rather than any item with similar or dissimilar features. All foils became targets in the Category Search task, and therefore were not coded separately in that task.
The participants completed 1650 trials throughout the experiment. For each trial, a search array displayed for 200 ms contained two black line-drawn aliens on a white background (Figure 3; see Eimer, 1996 and Wu et al., 2015 for examples of 2-item N2pc visual search). The duration was chosen to minimize eye-movements, which interfere with the N2pc ERP. Two stimuli were displayed instead of using a more complex search display with 4 or more items because the task was already difficult enough with two item arrays. The two aliens were displayed on each side of the fixation point (3.44° from center) to elicit the N2pc from contralateral and ipsilateral electrodes. Following the search array, a response screen displayed for 1600ms included only a black fixation dot (Figure 3). Participants were asked to fixate on the dot throughout the entire experiment and indicate target presence or absence with left and right arrow keys using the right hand.
We DC-recorded the EEG at standard positions of the extended 10/20 system (500 Hz sampling rate; 40 Hz low-pass filter) using 32 electrodes. The EEG was re-referenced offline to the averaged earlobes. The data were split into epochs from −100ms to 500ms relative to the search array onset, with a pre-stimulus baseline of 100ms. We used the following criteria for artifact rejection for the entire epoch: horizontal EOG exceeding ± 25 μV, vertical EOG exceeding ± 60 μV, all other channels exceeding ± 80 μV. Including only correct trials, the average percentage of trials retained per participant after artifact rejection was 74%, the typical amount in previous studies (e.g., Wu et al, 2015). Mean N2pc amplitudes were obtained at lateral posterior electrodes PO7 and PO8 between 220ms and 340ms after search array onset (Wu et al., 2013).
To investigate accuracy effects for target present trials (i.e., hits), a 2 (Intra-item feature similarity: Family S vs. D) x 2 (Trial type: Exemplar vs. Category match) repeated-measures ANOVA revealed a main effect of trial type (F(1,15)=215.19, p<.001, η2=.94) and a main effect of intra-item feature similarity (F(1,15)=26.51, p<.001, η2=.64) (Figure 4). Accuracy for Exemplar match trials was greater than that for Category match trials overall. In addition, accuracy was greater for items with similar features compared to dissimilar features for both Exemplar and Category match trials. There was no interaction between these variables for accuracy (F=.05).
An ANOVA for reaction time also revealed a main effect of trial type (F(1,15)=50.97, p<.001, η2=.77) and a main effect of intra-item feature similarity (F(1,15)=17.59, p<.001, η2=.54), and no interaction (F=.56). RTs were faster for items with similar intra-item features than for dissimilar features in both Exemplar and Category match trials, and when searching for one specific item compared to category search.
To investigate accuracy effects for target absent trials (i.e., correct rejections), a 2 (intra-item feature similarity: Family S vs. D) x 2 (Trial type: Foil vs. No Target) repeated-measures ANOVA revealed only a main effect of trial type (F(1,15)=79.77, p<.001, η2=.84) (Figure 4), where accuracy was higher for Foil trials compared to No Target trials.
An ANOVA for reaction time also revealed a main effect of trial type (F(1,15)=128.16, p<.001, η2=.90), where RTs for Foil trials were faster than RTs for No Target trials. There was also a main effect of intra-item feature similarity (F(1,15)=4.71, p=.05, η2=.24), where RTs were faster for similar compared to dissimilar items for both Foil and No Target trials.
Overall, these behavioral results show that both exemplar and category search benefited from intra-item feature similarity (despite inter-item dissimilarity across both categories), and searching for a specific object was easier than searching for a category of objects. The EEG results in the next section determined whether the observed behavioral benefits from intra-item feature similarity were due to establishing a categorical attentional template based on grouping similar features.
Planned t-tests were conducted to assess the presence of the N2pc component in all trial types (Exemplar, Foil, Category) for both similar and dissimilar aliens (adjusted α = .05/3 = .02 for conducting 3 pairwise comparisons for similar aliens, and 3 comparisons for dissimilar aliens). A significant N2pc component was found for trials where targets were either an exemplar from Family S (t(15)=-3.89, p=.001) or an exemplar from Family D (t(15)=-5.20, p<.001), as expected based on numerous prior findings at the exemplar-search level. Importantly, the N2pc was also significant for any alien (Category target) from Family S (t(15)=-4.78, p<.001) (Figures 5 and and6),6), but not any alien from Family D (t(15)=.72, p=.48). Finally, foil trials with either similar or dissimilar aliens failed to show a significant N2pc (both ts<1.34, ps>.20) (Figures 5 and and6),6), confirming that these novel items had to be members of the task set to generate an N2pc.
Given our interest in comparing target present trials (Exemplar and Category match) between categories and that foil trials for both similar or dissimilar categories did not produce N2pc components, we excluded foil trials from further analyses. Investigating the difference between similarity conditions for Exemplar and Category target trials, a 2 (laterality: contralateral vs. ipsilateral amplitude) x 2 (task: Exemplar vs. Category search) x 2 (Intra-item similarity vs. dissimilarity) repeated-measures ANOVA revealed a main effect of laterality (F(1,15) = 27.47, p<.001, η2=.65), and an interaction between laterality and task (F(1,15) = 18.00, p=.001, η2=.55), which indicated larger contralateral N2pc components for Exemplar than for Category match trials. Critically, there was a 3-way interaction between laterality, task, and intra-item feature similarity (F(1,15) = 5.17, p=.04, η2=.26). There were no other main effects or interactions (Fs< 3.59).
To investigate the 3-way interaction, two pairwise t-tests were conducted to compare the N2pc components for each trial type, based on feature similarity (adjusted α = .03). The N2pc amplitude of category search for items with similar features (i.e., any alien from Family S) was larger than that for Family D (t(15)=3.85, p=.002) (Figure 7). Fourteen out of 16 participants had larger N2pc components for the category search task with high intra-item similarity (Family S) compared to the category search task with low intra-item similarity (Family D), which did not produce a significant N2pc based on the analyses presented earlier. The N2pc amplitude for the Exemplar Match trials did not differ based on the presence or absence of high intra-item feature similarity (t(15)=.17, p=.86). Overall, these ERP results show that exemplar search did not benefit from intra-item feature similarity, while category search did benefit.
To determine how lower-level perceptual grouping facilitates attentional selection of higher-level novel categories, the present study investigated whether intra-item similarity facilitates categorization, despite inter-item dissimilarity. While the perceptual similarity literature (e.g., Duncan & Humphreys, 1989) would likely predict that inter-item dissimilarity hinders categorization, findings from research on consistent feature correlations (e.g., Austerweil & Griffiths, 2011; Younger & Cohen, 1986) suggest that intra-item similarity may facilitate categorization via grouping of category members. Using a visual search paradigm, the present study provides support for the feature correlation literature: Intra-item feature similarity not only facilitated exemplar search, but also category search, despite inter-item dissimilarity. Intra-item feature similarity led to higher accuracy and faster reaction times in both exemplar and category search tasks compared to search for aliens with dissimilar intra-item features.
To measure whether the search efficiency based on behavioral measures was modulated by the grouping of distinct attentional templates into an abstract categorical template, we measured the N2pc, the ERP marker for target selection emerging at approximately 200 ms. The N2pc is modulated by the number of items guiding search, unless the search is for an existing or recently trained category of items. During category search, the N2pc is present regardless of the number of items within the category (Nako et al., 2014a). We found that the category of items with similar intra-item features (and not the category with dissimilar features) elicited an N2pc, indicating that the ability to group features at the item level facilitated grouping at the category level. These ERP results are noteworthy because we used identical features across both categories; the only difference between categories was whether the features on each alien were paired or shuffled. Importantly, the aliens in the intra-item similarity category were perceptually dissimilar from each other, just as the items from the other category. While an N2pc-like component may have emerged at a later time window (400 ms) for the intra-item dissimilar category, the SPCN/CDA (working memory component, Eimer 2014) is also measured on the same electrodes as the N2pc and typically emerges at 400 ms. Therefore, it is difficult to differentiate between a late N2pc component and a typical SPCN/CDA component in our paradigm. Future work should determine whether a “late” N2pc can be interpreted in the same way as a typical N2pc.
Based on our N2pc results, the early grouping benefit of similar features was obtained only for category search, and not for exemplar search, suggesting that participants could bind features within one object into one unit regardless of feature similarity. While participants considered the eight items in the similar category as a unit, this grouping benefit did not interfere with exemplar search trials. There was no significant N2pc during foil trials to category-matching non-targets (e.g., circle back, circle belly) when searching for a specific target (e.g., square back, square belly). In other words, category-matching non-targets did not capture attention in an obligatory manner, as they do in well-learned categories (e.g., letters, numbers, clothing; Nako et al., 2014a; Nako et al., 2014b). Rather, feature similarity defined a circumscribed category of objects, and object categories differed from each other based on the presence or absence of feature similarity.
The discrepancy between the behavioral and neural results may be perplexing at first, namely no difference during Exemplar search in neural responses with a difference in behavioral responses. However, a number of studies document discrepancies between neural and behavioral results (Haynes & Rees, 2005; He et al., 1996). Recent N2pc research suggests that behavioral responses are modulated by additional factors that do not affect the N2pc component (Wu et al., 2015). In Wu et al., participants had slower reaction times when selecting non-native targets (ape faces with which we tend to have little experience) compared to native targets (human faces with which we have a lifetime of experience). This behavioral effect is entirely consistent with previous behavioral studies on perceptual narrowing, a developmental phenomenon where we learn about the people and languages in our environment at the cost of a decreased ability to process non-native images and sounds (c.f. Scott et al., 2007). However, Wu et al. (2015) found that the N2pc did not differ between native and non-native stimuli, even though behavioral responses differed significantly. Perhaps, similar to Wu et al. (2015), participants in the current study exhibited more hesitation to verify the presence of a target when searching for items with intra-item dissimilarity. The N2pc is a fast involuntary implicit response and perhaps immune to voluntary hesitations that are evident in behavior (e.g., guidance/verification phase, see Eimer, 2014). Future work should determine whether participants delay behavioral responses in this task due to an extended verification phase (see Maxfield & Zelinsky, 2012).
To support our claim that grouping multiple exemplar templates into a categorical template underlies our findings, we can rule out bottom-up attentional capture to feature similarity as an explanation for our ERP results for two main reasons. First, the N2pc components were very comparable in the exemplar trials for both similar and dissimilar aliens. Moreover, there was no N2pc component for foil trials (i.e., non-target aliens in the same category as the exemplar target). If lower-level bottom-up capture were solely responsible for our findings, we would have observed larger N2pc components within the similar category (compared to the dissimilar category) for both the exemplar and foil trials, as well as the category trials. Second, the exemplar and category search trials only differed in the top-down task set and explicit instructions, not in the actual stimuli displayed. The exemplar search trials had larger N2pc amplitudes compared to the category search trials. If bottom-up capture were the source of our N2pc effects, the N2pc components between the exemplar, foil, and category trials would have been identical.
One unresolved issue in this study is what comprises a categorical template. We propose that in this study it is based on visual features organized by a perceptual rule at the item level. Therefore, it is neither based solely on visual features (e.g., any green object) nor completely rule-based (e.g., things to bring on vacation). While making judgments based on matching features is a perceptual decision, identifying category targets in this study required a rule that bound all of the targets into one category (i.e., any alien with similar features). Participants could not merely rely on a perceptual template because inter-item similarity was low among category members. Many naturalistic categories also require the use of both visual features and rules (e.g., letters, numbers, clothing), though some may be more perceptually based (e.g., cars) than others (e.g., food) (see work on hybrid search, Cunningham & Wolfe, 2014). While real-world category-level search is much more complex than that targeted in our study, we used a well-controlled single case to better understand how category knowledge can be built easily with novel stimuli. Furthermore, category learning studies typically use a variety of category types, including both familiar and novel categories based on abstract or perceptual criteria (e.g., simple arbitrary feature combinations such as tail length and number of fingers on novel “bugs”, Sloutsky, 2010; see also Blair et al., 2009; Posner & Keele, 1968; Younger & Cohen, 1986). Previous work has shown that searching for homogenous perceptually-based categories that can be identified via diagnostic features (e.g., ape faces vs. other animal faces) deploys perceptual category-level templates by 200 milliseconds (Wu et al., 2015). It is not too surprising that searching for a specific item and searching for diagnostic features from a single category template across multiple items have similar response patterns. Other studies have shown that familiar abstract categories (e.g., clothing, kitchen items) elicit similar N2pc components to perceptually based categories (e.g., Nako et al., 2014b). The novel finding in the current study is that the use of newly acquired abstract category templates occurs within the same time window as perceptual templates even when there are no specific diagnostic perceptual features for a particular category. This finding, along with previous findings, shows that the N2pc ERP component is a robust marker of categorization for both familiar and novel homogenous (perceptual) and heterogeneous (abstract) categories. Recent work from our lab extends the current finding to show that the N2pc also emerges within the same time window for very broad heterogeneous categories such as “healthy vs. unhealthy food” (Wu et al., under review). While it is clear from our results that abstract category templates can be used within the same time window as perceptual templates, more research is required to understand how abstract category templates are constructed, stored, and deployed relative to perceptual templates.
One limitation of the current study is that we used neural evidence to infer a cognitive process: categorical template deployment. However, our hypotheses were generated based on extensive previous research. We measured a specific, well-documented ERP component that is a known marker of target selection, narrowing down the possible cognitive processes used in the task leading to the differences in the N2pc generated for the two item sets. In addition, we controlled for the previously mentioned factors (i.e., reward salience, number of presented targets, semantic relatedness between targets and distractors, and pop-out effects) that have been shown to modulate the N2pc independent of grouping mechanisms. In doing so, we reduced the risk of false alarms, by reducing the number of alternative explanations (Poldrack, 2006). There was one alternative explanation for our findings that we did not explore in this study due to visual search constraints. We chose to include 8 exemplars per category to equate the number of potential targets in each category. However, the actual potential number of targets in the “different” category is 256 compared to 8 aliens in “similar” category. Of note, previous studies show that actual set size does not modulate the N2pc after controlling for perceptual similarity across exemplars (Wu et al., 2013). In addition, two of our unpublished studies show that even broad categories such as “healthy food” or “upright items” elicit a reliable N2pc in the appropriate time window. Therefore, it is unlikely that the potential number of targets determines the presence or absence of the N2pc independent of categorization, though it may modulate the amplitude of the N2pc. Based on previous studies, our current finding provides strong evidence that under specific circumstances the N2pc can be a neural signature of rapid category-based selection, and that this can be elicited by intra-item similarity.
The current study adds to a growing body of work investigating how the process of grouping items reduces attention and working memory limitations (see Orhan et al., 2014 for review; Austerweil & Griffiths, 2011; Brady et al., 2009; Orban et al. 2008; Woodman et al., 2003). Grouping can be implemented via statistical regularities (e.g., Brady et al., 2009; Orhan et al., 2014), feature correlation (e.g., Austerweil & Griffiths, 2011; Wu et al., 2013), and Gestalt principles (e.g., Woodman et al., 2003). Moreover, grouping has been observed even in young infants (e.g., Quinn & Bhatt, 2009; Wu et al., 2011), highlighting the fundamental nature of this mechanism. The novelty of the current findings is that grouping at the lower item level facilitates grouping at the higher category level. This grouping benefit allows us to overcome efficiency limitations in visual search. In turn, visual search studies can reveal how items are grouped across various task demands, and search measures can be used as markers of learning and categorization (e.g., Wu et al., 2013).
Future work should further investigate how feature correlation (other than intra-item similarity) and other statistical regularities might facilitate attentional selection of novel categories (e.g., Wu et al., 2013). A large literature shows that regularities (correlation, co-occurrence, transitional probabilities) help infants and adults isolate and group features in both visual and auditory domains (grouping features into objects and words, e.g., Fiser & Aslin, 2001; 2002; Kirkham et al., 2002; Saffran et al., 1996; Wu et al., 2011). How such knowledge becomes useful for the learner in selecting future information is an important issue in developmental psychology and cognitive science. Investigations into how naïve and mature learners use such regularities to find grouped multi-part targets among distractors would allow for a better understanding of how we efficiently learn to attend.
We thank Kristen Meyer, Jared Band, Sandy Xiong, and Eric Partridge for their help with data collection. We also thank Robert Jacobs, George Malcolm, Melissa Vo, Marisa Carrasco, Paul Quinn, Benjamin Zinszer, Rebecca Nako, and UCR CALLA Lab RAs for useful discussions on the study design, data interpretations, and comments on earlier versions of the manuscript. We also thank Nicholas Turk-Browne and 2 anonymous reviewers for insightful comments on the manuscript. This research was conducted at the University of Rochester with funding from an NRSA grant (F32HD070537) from NICHD to R.W., and from an NIH grant (HD-037082) to R.N.A.