A visual short-term memory advantage for objects of expertise

Abstract

Visual short-term memory (VSTM) is limited, especially for complex objects. Its capacity, however, is greater for faces than for other objects, an advantage that may stem from the holistic nature of face processing. If the holistic processing explains this advantage, then object expertise—which also relies on holistic processing—should endow experts with a VSTM advantage. We compared VSTM for cars among car experts to that among car novices. Car experts, but not car novices, demonstrated a VSTM advantage similar to that for faces; this advantage was orientation-specific and was correlated with an individual's level of car expertise. Control experiments ruled out accounts based solely on verbal- or long-term memory representations. These findings suggest that the processing advantages afforded by visual expertise result in domain-specific increases in VSTM capacity, perhaps by allowing experts to maximize the use of an inherently limited VSTM system.

Introduction

Each and every one of our interactions with the world is constrained by narrow bottlenecks of information processing, including how many pieces of visual information we can retain in memory. Is there anything a person can do to increase his or her visual short-term memory (VSTM) capacity? Typically, people are only able to retain three to four objects in VSTM at any given time. One possibility is that VSTM capacity is determined by a fixed number of “slots” (3-4) that can hold one object each (Vogel, Woodman, & Luck, 2001). However, some have pointed out that it may be limited by the complexity or number of features of the items stored (Alvarez & Cavanagh, 2004; Wheeler & Treisman, 2002). One possibility is that visual expertise can help an observer overcome such limitations; visual experts process highly complex objects within their domain of expertise with relative ease, creating qualitatively different “holistic” representations (Gauthier, Curran, Curby, & Collins, 2003; Tanaka & Sengco, 1997) that support faster identification judgments (Tanaka, 2001; Tanaka & Taylor, 1991) and can be searched through more efficiently (Tong & Nakayama, 1999). But do such advantages impact VSTM capacity?

Recently, we showed that VSTM capacity for upright faces is larger than that for other categories, such as cars, watches or even inverted faces (Curby & Gauthier, 2007). One possibility is that humans are innately endowed with a greater memory capacity for upright faces due to the importance of face memory for survival. Alternatively, superior VSTM capacity for upright faces may be a product of our expertise with this category.

Currently, there is limited support for an impact of experience on VSTM capacity. Some evidence suggests that VSTM capacity remains stable from early in development (12 months; Rose, Feldman, & Jankowski, 2001) to adulthood (Luck & Vogel, 1997) suggesting it may be relatively inflexible. However, more recent studies have found that children's VSTM capacity for simple colored shapes doubles throughout childhood, from 2 items at five years of age to the adult-like capacity of 3-4 items by ten years of age (Cowan et al., 2005; Riggs, McTaggart, Simpson, & Freeman, 2006). Similar increases in capacity with development have also been reported for verbal short-term memory (Cowan, Nugent, Elliott, Ponomarev, & Saults, 1999). At the moment, it is unclear which aspects of cognitive development could be at the basis of such dramatic changes.

Other studies have employed training regimens to investigate whether VSTM capacity is influenced by experience, and they have generally found little or no training effect. For example, a recent study exploring the influence of domain-specific training on VSTM for novel objects reported that participants who viewed a set of eight random polygons 160 times in the context of a VSTM task were no more accurate detecting a change in a VSTM array containing these trained shapes compared to unfamiliar ones (Chen, Eng, & Jiang, 2006). Importantly, participants in this study could accurately identify the trained polygons in a two-alternative-forced-choice task, confirming that they did have representations of these items in long-term memory (LTM). Notably, VSTM performance did improve with practice, but equally for both trained and untrained polygons. Thus, it is unclear whether this change in performances represents a general effect of practice on VSTM performance, or a more specific influence of experience that generalizes to new exemplars within the trained category. It is also possible that more extensive training, such as that required to develop perceptual expertise with a category of objects, is required to induce a change in VSTM capacity beyond that attributed more generally to practice.

A recent neuroimaging study adopted a more extensive (10.5 hour) training paradigm to explore the impact of experience on the neural substrates supporting VSTM (Moore, Cohen, & Ranganath, 2006). Training increased activity during both encoding and maintenance of artificial objects in the classic VSTM network, including the bilateral dorsolateral prefrontal, posterior parietal, and occipitotemporal cortices. In contrast, the lateral occipital (LO) cortex and the fusiform face area (FFA) showed expertise effects during encoding only. These changes in the functional network supporting VSTM for trained stimuli observed after extensive training are consistent with the suggestion that behavioral changes in VSTM capacity may be possible after extensive training. However, VSTM capacity was not measured in this study, so it is difficult to know whether increases in capacity would correlate more with changes in encoding or maintenance.

Some might suggest that the superior VSTM documented in chess experts, relative to chess novices, for configurations of chess pieces can be considered evidence that experience can increase the capacity of VSTM (Chase & Simon, 1973). However, this advantage is believed to rely on specific stored representations in long-term memory, rather than a more qualitative change in the way information is stored in VSTM (Chase & Simon, 1973). Support for this LTM account of expert chess memory comes from the fact that intervening short-term memory tasks during the retention interval do not impact memory performance for familiar chess positions (Charness, 1976). Practice appears to increase chess experts' VSTM by allowing “chunking” of information into larger units in long-term memory and storing pointers to these chunks in VSTM (Chase & Simon, 1973; Freyhof, Gruber, & Ziegler, 1992; Gobet & Simon, 1998). Therefore, although previous studies have demonstrated a benefit of experience or expertise on VSTM capacity, these advantages have been shown to reflect the utilization of additional resources, such as LTM or verbal memory, to supplement VSTM rather than a change to VSTM capacity (Charness, 1976; Chase & Ericsson, 1981).

Here, we are concerned with a different way in which experience may influence VSTM. Specifically, the impact of the perceptual organization of information on VSTM capacity (Delvenne & Bruyer, 2004; Luck & Vogel, 1997; Vogel et al., 2001; Xu, 2002; Xu, 2006) may provide a potential avenue for extensive learning to influence VSTM capacity. For example, VSTM capacity is greater for features presented in the form of a unified object, rather than for those presented in isolation (Luck & Vogel, 1997). Thus, in addition to recruiting additional capacity from other systems, as in the case of chess experts, experience may impact VSTM capacity more directly because of a change in the manner in which an item is encoded and/or represented in VSTM. For example, faces and other objects of expertise are processed more holistically compared to objects of non-expertise, which are processed in a more feature-based manner (Busey & Vanderkolk, 2005; Gauthier et al., 2003; Gauthier & Tarr, 2002). Classic holistic processing effects typically found with faces, such as sensitivity to inversion or difficulty selectively attending to an object part presented in the context of a whole object, have been demonstrated among observers trained to become experts with a novel category (Gauthier & Tarr, 2002) and also among real-world car experts (Gauthier et al., 2003). Thus, expertise with faces and also chess appears to lead to the processing and/or storage of information in larger units or chunks. However, unlike the memory strategy used by chess experts that operates over multiple independent meaningful pieces with specific meaning associated to different arrangement of these pieces, holistic processing of faces and objects operates within an item (e.g. a car), not unlike other object-based perceptual advantages previously reported in the literature (e.g., Egly, Driver, Rafal, 1994; Saiki & Hummel, 1998; Xu, 2006). Notably, holistic processing has been related to activity in the FFA (Gauthier & Tarr, 2002; Rotshtein, Geng, Driver, & Dolan, 2007), which is engaged during the encoding of faces (Druzgal & D'Esposito, 2001; 2003) and objects of expertise (Moore et al., 2006). This processing strategy has also been associated with the earliest face-specific electrophysiological potential, the N170, occurring only 170 ms after the onset of a stimulus (Gauthier et al., 2003). Therefore, it is possible that in the case of faces, rather than relying on LTM, the VSTM advantage may stem from differences at a more perceptual level.

In sum, although there is a documented VSTM advantage for upright faces (Curby & Gauthier, 2007) and a literature suggesting that most observers are experts with upright but not inverted faces, this can only indirectly support inferences about the effect of perceptual expertise on VSTM capacity. In the current studies we specifically set out to test if expertise can increase VSTM capacity by comparing experts and novices with a non-face category: cars. Previous studies have found that cars are processed more holistically among car experts than car novices (Gauthier et al., 2003) and that a subject's car expertise is related to the amount of activity in response to cars in the part of the visual system most responsive to faces, the fusiform face area (Gauthier et al., 2000, 2005; Xu, 2005). Notably, activity in this area has also been linked with holistic processing (Gauthier & Tarr, 2002; Rotshtein et al., 2007).

Experiment 1

Experiment 1 assessed VSTM capacity for upright and inverted faces and cars among participants with a range of perceptual expertise with cars. Given our hypothesis that holistic processing underlies the VSTM advantage for faces and because inversion disrupts holistic processing (Tanaka & Sengco, 1997), we predicted an expert VSTM advantage for upright, but not inverted, faces and cars. In addition, we varied encoding time up to 4000-ms to ensure VSTM was not limited by encoding speed, as complex objects require more time to be encoded in VSTM than do simple objects (Curby & Gauthier, 2007; Eng, Chen, & Jiang, 2005).

Methods

Participants

Thirty-six individuals ranging in experience identifying cars participated for payment. Participants were employees, undergraduate students, or graduate students of Vanderbilt University, or members of the surrounding Nashville community. All had normal or corrected to normal vision. A self-report measure of participants' car and bird expertise was obtained in the form of a rating on a scale of one to ten. Participants were informed that “five” corresponded to average skill at identifying cars or birds whereas “ten” reflected perfect skill recognizing these categories. An objective measure of car expertise was also obtained using a sequential matching task used in previous studies (Gauthier, Skudlarski, Gore, & Anderson, 2000; Gauthier et al., 2003, 2005). In this task, participants were required to make a same/different judgment about different images of cars at the level of model, regardless of year (see Figure 1A in Gauthier et al., 2003). This task can be performed at least to some minimal degree by all participants, regardless of their level of experience with cars, as it does not require knowledge of car names. To provide a baseline of their perceptual skills, participants also performed the same task with birds, in which they were required to make a same/different decision at the level of species about different images of passerine birds. A car expertise index was defined as (car d′–bird d′). Participants with a car expertise index ≥1 and a d′ for cars ≥2 were classified as experts (Gauthier et al., 2000).

Figure 1.

The sequence of events in each trial: Participants were first presented auditorily with two digits and a mask, which they overtly rehearsed throughout the trial to prevent verbal rehearsal. The study array then appeared for 500-, 2500-, or 4000-ms. After a 1200-ms delay a face or car probe was presented in one of the locations from the study array. The probe remained until participants indicated with a key press whether the probe was the same as or different from the one that appeared in that location in the study array. After a response was made, a screen with two digits appeared and participants were required to state whether the two digits on the screen are the same as those they had been rehearsing throughout the trial.

Eighteen participants (11 males) met the criteria for car expertise (age, M=22.28, SD=4.71, car d′ M=2.55, bird d′ M=0.87), while the remaining 18 (10 male) were classified as car novices (age, M=20.64, SD=2.42, car d′ M=1.34, bird d′ M=0.84)¹. One of the participants classified as an expert had a car d′ of 2.24 but a (car d′ – bird d′) less than one (0.83); he was included in the car expertise group because he also reported having above average skills at recognizing birds, which likely resulted in the smaller difference between the d′ prime measures for these categories. Car expertise scores from the matching task were generally consistent with subjects' self-report, with participants classified as novices reporting their car recognition skills at an average of 5.77/10; those who met criteria for car expertise rated their skills, on average, as 8.00/10.

Stimuli

The stimuli were 72 grayscale faces (1.9°×2.3°) from the Max-Planck Institute for Biological Cybernetics in Tuebingen, Germany (Troje & Bülthoff, 1996), and 72 grayscale images of cars, gathered from various public web sites (2.3°×1.5°, profile view).

Procedure

For each participant, half the faces and half the cars appeared in the upright trials. The remaining appeared in the inverted trials. Participants performed a delayed match-to-sample probe recognition task simultaneously with an articulatory suppression task (Figure 1). The sequence of events in each trial was as follows: participants were first presented aurally with two digits and a mask, which they overtly rehearsed throughout the trial to prevent verbal rehearsal. The study array, consisting of 1, 3, or 5 faces or cars evenly spaced in a circle (6.1° diameter) (either all upright or all inverted), then appeared for 500-, 2500-, or 4000-ms. After a 1200-ms delay a face or car probe was presented in one of the locations from the study array. The probe remained until participants indicated with a key press whether the probe was the same as (50% of trials) or different from the one that appeared in that location in the study array. To minimize confusion, within each trial the probe was never an item that had appeared at a different location in the study array. After a response was made, a screen with two digits appeared and participants were required to state whether the two digits on the screen are the same as those they had been rehearsing throughout the trial.

Participants performed a total of 1152 trials across four different sessions, each consisting of 8 alternating blocks of upright and inverted images (36 trials/block, randomized for set-size and presentation duration). Two sessions consisted of only face trials, while the other two sessions consisted of only car trials. Session order was counterbalanced within and across expertise groups. In sum, there were 288 trials for each of the four categories (upright faces, inverted faces, upright cars, inverted cars). For each category, there were 9 conditions (3 set sizes × 3 durations), presented 32 times each.

Analysis

Incorrect articulatory suppression trials (< 2%) were discarded. For each participant and condition, the number of objects successfully encoded in VSTM was estimated using Cowan's K, where K= (hit rate + correct rejection rate - 1) * set size (Cowan, 2001). The maximum K (K-max) was identified for each duration regardless of set size. All analyses were performed on the K-max values. In addition to ANOVA and regression analyses exploring the relationship between level of car expertise and VSTM capacity, a series of planned t-tests were conducted to explore the specific predictions based on the proposed role of holistic processing in increasing VSTM capacity, that is (1) the presence of an inversion cost for cars among car experts but not novices, (2) greater VSTM for cars among car experts compared to novices, and (3) greater VSTM capacity for faces than cars among car novices, with sufficient encoding time.

Results

Both car experts and car novices demonstrated an inversion cost for faces, while only car experts experienced such a cost for cars (Figure 2A, 2B). Car experts also demonstrated greater VSTM for upright cars than novices when the presentation duration was sufficiently long (≥ 2500-ms). Furthermore, while VSTM for faces was not different from that for cars among car experts regardless of presentation duration, car novices demonstrated an advantage for upright faces over upright cars at the longest presentation (4000-ms). With a 4000-ms presentation, car expertise was correlated with VSTM for cars but not for faces (Figure 2C, 2D).

Figure 2.

The maximum number of objects (K-max) in visual short-term memory (VSTM) for 500-, 2500-, and 4000-ms presentation durations for (A) upright and inverted faces and (B) upright and inverted cars among participants who were car experts and novices. There was a VSTM advantage for upright cars among cars experts similar in magnitude to that for upright faces. Car experts, but not novices, showed an inversion effect for cars. Error bars represent pooled standard error values. Scatter plots of individuals' car expertise scores (Car d′–Bird d′) and their K-max when the memory array was presented for 4000-ms illustrate the significant correlation between a participant's level of car expertise and VSTM capacity for (C) upright cars but not (D) upright faces.

VSTM for faces among car experts and novices

For faces, a 2 (orientation: upright, inverted) × 3 (duration; 500-, 2500-, 4000-ms) × 2 (group; novice, expert) ANOVA on K-max revealed main effects of orientation, F(1,34)=69.53, p≤.0001, and duration, F(2,68)=40.24, p≤.0001, but no main effect (F<1) or interaction involving car expertise (all ps>.425). The interaction between orientation and duration failed to reach significance, F(2,68)=1.77, p=.178. In sum, face VSTM was greater for longer presentations and for upright than inverted faces, but was not influenced by car expertise.

VSTM for cars among car experts and novices

For cars, a 2 (orientation: upright, inverted) × 3 (duration; 500-, 2500-, 4000-ms) × 2 (group; novice, expert) ANOVA on K-max revealed main effects of orientation, F(1,34)=9.12, p=.0048, and duration, F(1,34)=61.93, p≤.0001, but no interaction between orientation and duration (F<1). Although there was no main effect of car expertise, F(1,34)=1.95, p=.172, there was an interaction between expertise and orientation, F(1,34)=5.61, p=.024. Interactions between duration and expertise and/or orientation failed to reach significance, (all p's ≥.230). In sum, VSTM was generally greater for longer presentations and car experts showed superior VSTM for upright, but not inverted, cars relative to novices.

Planned comparisons

K-max for cars was greater for upright compared to inverted orientations for all durations among experts (all ps<.022) but not novices (all ps>.206), while this advantage for upright orientations existed for faces for all durations among both groups (experts, all ps<.010; novices, all ps<.033). In addition, with sufficient presentation duration, K-max for upright faces reliably exceeded that for upright cars among car novices (500-ms, t<1; 2500-ms, t(17)=1.48, p=.157; 4000-ms, t(17)=2.87, p=.011), but not experts (500-ms, t(17)=1.19, p=.251; 2500-ms, t(17)=1.16, p=.263; 4000-ms, t(17)=1.17, p=.257). A VSTM advantage for cars among car experts compared to novices emerged when the presentation duration was sufficiently long (500-ms, t<1; 2500-ms, t(34)=1.68, p=.103; 4000-ms, t(34)=3.05, p=.004).

Correlation between car expertise and VSTM for cars and faces

With a 4000-ms presentation, K-max for upright cars was correlated with participants' car expertise index (r=.364, p=.032). Notably, car expertise indexes were not correlated with VSTM for upright (r=.032, p=.885) or inverted faces (r=.000, p=.951) or inverted cars (r=.036, p=.838)².

Discussion

Consistent with the proposed influence of perceptual expertise on VSTM capacity, car experts demonstrated an orientation-dependent VSTM advantage for cars. Similar to the VSTM advantage for faces, this advantage depended on sufficient encoding time (Curby & Gauthier, 2007). These results suggest that the VSTM advantage for faces is not due to a face-specific mechanism; other objects within a domain of expertise can also demonstrate this advantage.

The expert advantage reported here may stem, as we predicted, from differences in perceptual processing, but one alternative is that experts benefited from better knowledge of car names, possibly leading to a contribution from verbal short-term memory (Olsson & Poom, 2005). Experiment 2 explores this hypothesis.

Experiment 2

The articulatory suppression load used in Experiment 1 may have been insufficient to prevent a contribution from verbal short-term memory; participants can perform a verbal memory task with reasonable accuracy (82%) with an articulatory suppression load equivalent to the two syllables used in Experiment 1, but performance drops considerably (54%) when the load is increased to six syllables (Marsh & Hicks, 1998)³. Thus we increased the articulatory suppression load to 5-6 syllables in Experiment 2. A semantically relevant load was also used to further interfere with any verbal rehearsal strategy; participants were required to rehearse three car model names during car trials and three person names during face trials.

Methods

Participants

Thirty-one individuals, whose car expertise was quantified as in Experiment 1, participated for payment. Fourteen participants (11 males) met the criteria for car expertise (age, M=21.64, SD=2.10, car d′ M=2.72, bird d′ M=0.93), while 17 (13 male) were classified as car novices (age, M=22.41, SD=3.02, car d′ M=1.27, bird d′ M=0.76)⁴.

Stimuli, Procedure & Analyses

The stimuli, procedure and data analysis were as in Experiment 1, but instead of rehearsing digits, participants rehearsed three car models (e.g. “Spectra, Blazer, Accord”; no overlap with models from visual task) or person names (e.g. “Leanne, Amy, Cathryn”). These were probed auditorily at the end of each trial.

Results

As in Experiment 1, both car experts and novices demonstrated an inversion cost for faces, while only car experts experienced such a cost for cars (Figure 3A, 3B). Furthermore, while VSTM for faces was no different from that for cars among car experts regardless of presentation duration, car novices demonstrated an advantage for upright faces over upright cars when the stimulus presentation was sufficiently long (4000-ms). In addition, car experts demonstrated greater VSTM for upright cars than novices only when presentation was sufficiently long (≥2500-ms). With 4000-ms presentation, car expertise was correlated with VSTM for cars but not faces (Figure 3C, 3D).

Figure 3.

The maximum number of items (K-max) in visual short-term memory (VSTM) for 500-, 2500-, and 4000-ms presentation durations for (A) upright and inverted faces and (B) upright and inverted cars among participants who were car experts and car novices. There was a VSTM advantage for upright cars among cars experts similar in magnitude as the advantage for upright faces. Car experts, but not novices, showed an inversion effect for cars. Scatter plots of individuals' car expertise scores (Car d′–Bird d′) and their K-max when the memory array was presented for 4000-ms illustrate the significant correlation between a participant's level of car expertise and VSTM capacity for (C) upright cars but not (D) upright faces.

VSTM for faces among car experts and novices

For faces, a 2 (orientation: upright, inverted) × 3 (duration; 500-, 2500-, 4000-ms) × 2 (group; novice, expert) ANOVA on K-max revealed main effects of orientation, F(1,29)=32.61, p≤.0001, and duration, F(2,58)=40.79, p≤.0001, but no main effect, F(1, 29)=2.12, p=.156, or interaction involving car expertise (Fs<1). There was no reliable interaction between orientation and duration (F<1). In sum, VSTM for faces was greater for longer presentations and for upright faces, but car expertise did not impact VSTM for faces.

VSTM for cars among car experts and novices

For cars, a 2 (orientation: upright, inverted) × 3 (duration; 500-, 2500-, 4000-ms) × 2 (group; novice, expert) ANOVA on K-max revealed main effects of orientation, F(1,29)=11.24, p=.002, and duration, F(1,29)=33.59, p≤.0001, but not car expertise, F(1,29)=2.71, p=.110. The interaction between orientation and duration was not reliable, F(2,58)=1.10, p=.340, but there were interactions between car expertise and orientation, F(1,29)=7.40 , p=.011, and car expertise and duration, F(1,29)=3.64, p=.032: car experts demonstrated greater VSTM for upright, but not inverted, cars compared to car novices, and they also benefited more from additional encoding time compared to novices. The interaction between duration, group and orientation was not reliable (F<1).

Planned comparisons

Our predictions were confirmed by planned 1-tailed t-tests: K-max for cars among car experts showed an advantage for upright over inverted orientations for presentations of 2500-ms or longer; 500-ms (t<1), 2500-ms, t(13)=3.47, p=.002, 4000-ms, t(13)=2.52, p=.013. In contrast, among car novices inversion failed to impact VSTM capacity for cars regardless of duration (all ts<1). VSTM capacity was greater for upright than inverted faces for all durations among both experts (all ps<.05) and novices (all ps<.02). Furthermore, K-max for upright faces reliably exceeded that for upright cars among car novices at the longest encoding duration (500-ms, t<1; 2500 -ms, t(16)=1.27, p=.112; 4000-ms, t(16)=2.09, p=.027). However, VSTM for cars and faces did not differ among car experts regardless of duration (500-ms, t<1; 2500-ms, t(13)=1.25, p=.117; 4000-ms, t<1). A VSTM advantage for cars among car experts compared to car novices only emerged with presentations longer than 500 ms (500-ms, t<1; 2500-ms, t(29)=2.42, p=.011; 4000-ms, t(29)=3.90, p=.0003).

Correlation between car expertise and maximum VSTM for cars and faces

K-max for upright cars with a 4000-ms presentation was correlated with participants' car expertise (r=.575, p=.0007). Notably, car expertise was not correlated with VSTM for inverted cars (r=-.027, p=.891)⁵, or upright (r=.000, p=.996) or inverted faces (r=.184, p=.322).

Discussion

Despite an increase in the articulatory suppression load, the VSTM advantage for faces and other objects of expertise remained intact; car experts, when given sufficient encoding time, not only demonstrated greater VSTM capacity, but also a greater inversion cost for cars than did novices. Similarly, novices demonstrated greater VSTM for faces than cars when given sufficient encoding time. Overall performance was also similar across Experiments 1 and 2. These results suggest that the greater VSTM capacity for cars among car experts does not rely on a contribution from verbal memory. It is possible that the knowledge of a label for a stimulus may change the manner in which it is processed in the VSTM task regardless of whether or not the label is explicitly accessed or used to aid recall. However, assuming a common underlying cause for the VSTM advantage demonstrated for faces and for cars among car experts, as suggested by the similar qualitative and quantitative nature of these two effects, the presence of this advantage for unfamiliar faces with no known labels provides evidence against this account (e.g. Experiments 1 & 2; see also Curby & Gauthier, 2007).

While the expert VSTM advantage does not appear to depend on a contribution from verbal short-term memory, it is possible that experts are better able to recruit or establish representations in LTM to aid VSTM performance. Experiment 3 explores this possibility.

Experiment 3

Among chess experts, LTM has been shown to play an important role in their superior ability to remember/recall meaningful configurations of chess pieces (Chase & Simon, 1973). Chess experts store large chunks of information about the spatial configuration of items in long-term memory that are recalled through a simple cue stored in VSTM. Similarly, it is possible that car experts' VSTM advantage may also depend on stimulus-specific representations in LTM (Gobet & Simon, 1998).

Visual expertise is clearly an example of long-term learning and it is reasonable to argue that any task recruiting such expertise must rely on at least some form of LTM. Experiment 3 explores whether the expert VSTM advantage depends on stimulus-specific LTM representations (Ericsson & Kintsch, 1995). Specifically, it is possible that the large stimulus sets (72 items/category) used in Experiments 1 and 2 allowed participants to use information in LTM: each item appeared infrequently (approximately 1:10) and the familiarity of true probes (relative to foils) could serve as useful cues to aid performance. Furthermore, experts' superior ability to distinguish exemplars might increase the reliability of familiarity cues. In Experiment 3, we used a small stimulus set in order to increase the frequency of item repetition, thus reducing the usefulness of LTM traces through the build-up of proactive interference. We additionally incorporated a manipulation check by changing the stimulus set partway through the experiment: if participants can still use LTM traces despite increased proactive interference, then their performance should drop when the stimulus set change occurs. Thus, Experiment 3 explored the potential role of stimulus-specific representations in LTM in contributing to the VSTM advantage for objects of expertise.

Methods

Participants

Thirty-six participants, whose car expertise was quantified as in Experiment 1, participated for payment. Eighteen (13 male) of which met our criteria for car expertise (age M=25.3, SD=4.54, car d′ M=2.84, bird d′ M=1.02), with the remaining 18 (8 male) classified as novices (age M=27.2, SD=8.36, car d′ M=0.70, bird d′ M=0.83)⁶.

Stimuli

The stimuli were 40 grayscale images of faces (1.9°×2.3°) and 40 profile views of cars (2.3°×1.5°).

Design, Procedure and Analysis

The design and procedure were similar to those used in Experiment 2, except that the set size was fixed at 5 and all images were upright and presented for 500-ms or 4000-ms. Participants performed 6 blocks of 36 trials for each category (faces, cars), totaling 432 trials. Trials for each category were performed in two separate (216 trial) sessions, with order of sessions counterbalanced across expert and novice groups. Twenty faces and 20 cars were randomly selected for each participant. Stimuli presented in the first 4 blocks for each category were selected from a subset of ten images. After 4 blocks of trials the stimulus set was switched to the remaining ten images. Each item appeared on average in every second trial.

Results

Neither a switch in stimulus set partway through the experiment nor the smaller size of the stimulus set eliminated the expert VSTM advantage (Figure 4A). Once again, a short encoding duration (500-ms) yielded relatively low VSTM capacity, which was not qualified by category or expertise effects. In the 4000-ms stimulus duration condition, only performance for cars among car novices was reduced by the stimulus set switch. VSTM capacity for faces exceeded that for cars among novices, but not among experts, for both pre- and post-switch stimulus sets. In addition, the VSTM advantage for cars among car experts, compared to novices, was apparent in the post-switch, but not pre-switch, stimulus set condition. Similarly, the correlation between car expertise and VSTM for cars was reliable for the post-switch, but not pre-switch, condition (Figure 4 B, 4 C).

Figure 4.

(A) The maximum number of upright faces or cars (K-max) in visual short-term memory (VSTM) before and after a stimulus set change. Participants had performed 140 trials with the same small (10 item) stimulus set before the stimulus set was switched and they performed 70 additional trials. Participants were either car experts or novices and the memory array was presented for 500- or 4000-ms. Only novice VSTM performance with cars in the 4000-ms presentation duration condition was influenced by the change in stimulus set. Scatter plots of individuals' car expertise scores (Car d′ – Bird d′) and their K-max for cars (B) pre- and (C) post- stimulus set switch. There was a significant correlation between a participant's level of car expertise and their VSTM capacity for cars in the post but not pre- stimulus switch condition.

VSTM for faces among car experts and novices

For faces, a 2 (duration; 500-ms, 4000-ms) × 2 (stimulus set; pre-switch, post-switch) × 2 (group; expert, novice) ANOVA revealed a main effect of duration, F(1,34)=135.50, p≤.0001, but no effect of stimulus set (F<1) or car expertise, F(1,34)=1.00, p=.324. In addition, no interactions between duration and/or expertise and/or stimulus set were reliable (all ps>.196). In sum, VSTM for faces was greater for longer presentations, but neither car expertise nor the change in stimulus set affected VSTM for faces.

VSTM for cars among car experts and novices

For cars, a 2 (duration; 500-ms, 4000-ms) × 2 (stimulus set; pre-switch, post-switch) × 2 (group; expert, novice) ANOVA revealed main effects of duration, F(1,34)=73.39, p≤.0001, and car expertise, F(1,34)=4.30, p=.046, but not stimulus set (F<1). In addition, there was an interaction between duration and expertise, F(1,34)=8.19, p=.007, with car experts only demonstrating a VSTM advantage for cars for the long presentation. The interaction between duration, expertise and stimulus set approached reliability, F(1,34)=3.37, p=.075 with novices' VSTM performance dropping after the switch in stimulus set but experts showing a slight increase.

Planned comparisons

Planned two-tailed t-tests explored the effect of the switch in stimulus set on VSTM capacity. Among car experts, capacity was unaffected by the switch regardless of duration for both cars (ts<1), and faces, 500-ms, t(17)=1.28, p=.218, 4000-ms, t<1. Among novices, VSTM performance was only reduced for cars at the longest duration, 4000-ms, t(17)=2.16, p=.045, 500-ms, t<1, with face VSTM performance unaffected by the switch regardless of duration, 4000-ms, t<1, 500-ms, t(17)=1.21, p=.241.

In addition, planned one-tailed t-tests revealed that VSTM for upright faces among car novices exceeded that for upright cars, as long as the presentation was sufficiently long (4000-ms), regardless of stimulus set (pre-switch, 500-ms, t(17)=1.24 p=.883; 4000-ms, t(17)=2.08, p=.027; post-switch, 500-ms, t<1; 4000-ms, t(17)=3.18, p=.0027). Among car experts, VSTM for cars and faces did not differ for either stimulus set regardless of duration (pre-switch, 500-ms, t(17)=1.32, p=.103; 4000-ms, t<1; post-switch, 500-ms, t<1; 4000-ms, t<1). A VSTM advantage for cars among car experts compared to car novices only emerged in the 4000-ms presentation condition for the post-switch stimulus set (500-ms, t<1; 4000-ms, t(34)=3.50, p=.0007). This car advantage in car experts failed to reach significance regardless of duration for the pre-switch stimulus set despite a trend in the long presentation condition (500-ms, t<1; 4000-ms, t(34)=1.29, p=.103).

Correlation between car expertise and VSTM for cars and faces

K-max for cars when the memory array was presented for 4000-ms reliably correlated with car expertise, but only for the post-switch stimulus set (pre-switch, r=.202, p=.236, post-switch, r=.452, p=.006). Notably, car expertise did not correlate with VSTM capacity for faces for either stimulus sets (pre-switch, r=.028, p=.870, post-switch, r=.074, p=.667).

Discussion

Neither the face advantage nor the expertise advantage for cars among car experts was eliminated by the build-up of proactive interference in LTM, which should have increased due to the small stimulus set in Experiment 3. In addition, consistent with the hypothesis that experts were not using stimulus-specific representations in LTM to aid their VSTM performance, we found no detectable cost to expert VSTM capacity when the stimulus set was replaced by an entirely different set of items after a period of learning. These results suggest that the expert VSTM advantage does not depend on access to stimulus-specific representations in LTM.

In contrast with the results from Experiment 1 and 2, where a larger stimulus set was used, VSTM capacity for cars in experts was not significantly higher than that for novices in the pre-switch condition (although there was a trend for such an effect). The use of a small stimulus set in Experiment 3 may have facilitated novices' performance; novices typically use a feature-based strategy to distinguish items (e.g. relying on the length of the trunk or the angle of the windscreen on a car), and thus the use of a small set of items in Experiment 3 would make such features even more diagnostic, allowing novices to adequately represent and distinguish items in VSTM. The success of such a strategy in novices would render experts' advantage for encoding highly complex objects moot.

The drop in VSTM for cars among car novices after a change in stimulus set could reflect the relative inflexibility of feature-based strategies in which the cars are identified by a single salient feature. After all, the relative usefulness of a feature as a distinguishing characteristic would critically depend on the variability of that feature among the items in the stimulus set. A feature-based strategy could be quite effective for a small stimulus set, consistent with the high-level of performance in the pre-change condition among novices. But this strategy would presumably be suboptimal for transfer to a different set of objects, as the same features are unlikely to be diagnostic across stimulus sets. In contrast, the more holistic perceptual strategy believed to be recruited for faces by both car experts and car novices, and for cars among car experts, would transfer equally well to a new stimulus set. Therefore, the drop in car VSTM performance among novices after the stimulus set switch may reflect different encoding strategies used by novices and experts.

The presence of the expert VSTM advantage under conditions where there should have been significant proactive interference in LTM provides additional evidence against the reliance of this effect on stimulus-specific representations in LTM, although it has been suggested that memory experts, such as digit span experts, may be able to overcome the influence of proactive interference by employing one of two strategies (Ericsson & Kintsch, 1995). For example, the most recently stored item can be distinguished based on its temporal context. However, it is unlikely that this temporal information would be sensitive enough to be reliable under conditions such as those in Experiment 3, where items frequently appeared in consecutive trials (a little more than a few seconds apart at times) and subjects performed 144 of such trials, with six faces per trial, within a half hour period. Alternatively, Ericsson and Kintsch (1995) suggested that experts can minimize proactive interference by generating multiple unique meaningful associations for the same chunk of information. Once again, while the digits typically used in Ericsson's studies can be easily encoded as a running time, a zip code, or a birthday, it would be considerably more difficult to implement such reliable alternative encoding strategies for differentiating unfamiliar faces at the individual level. Recall that in Experiment 3, each stimulus repeats up to 80 times, and thus a large number of different reliable alternative encoding strategies would have to be generated in order to avoid proactive interference. Thus, the idea that the VSTM expert advantage relies on accessing stimulus representations in LTM finds little or no support in Experiment 3.

Some might suggest that the VSTM advantage among experts may not have been affected by a change in stimulus set because they already had established LTM representations for the cars in both the pre- and post-change stimulus sets. In contrast, novices—who were presumably less familiar with the cars—may have been influenced by the additional exposure to the cars in the pre- change set as it may have provided an opportunity to establish item-specific representations in LTM. However, this account would have predicted that VSTM for unfamiliar faces would have incurred a cost due to the change in stimulus set, as participants could not have had pre-existing representation of these faces in LTM. In addition, this explanation would have also predicted that novices' VSTM for cars, and also VSTM for the unfamiliar faces more generally, would have increased from the first to the second half of the pre-change trials, which it did not⁷. Therefore, the use of pre-existing representations in LTM by experts, whether of faces or cars, appears to be unable to account for the pattern of results found in Experiment 3. In Experiment 4, we more directly evaluate the potential role of familiarity in contributing to the VSTM advantage for objects of expertise.

Experiment 4

Although the findings of Experiment 3 provide strong evidence that the VSTM advantage for objects of expertise does not depend on stimulus-specific LTM representations acquired during the study, it may be that further evidence is necessary before ruling out contributions from long-term stimulus familiarity. Specifically, it is possible that experts might still be able to recruit pre-existing representations in LTM acquired through their extensive real-world experience. One way to test this hypothesis would be through a comparison of VSTM performance for familiar and unfamiliar cars. However, comparisons using less familiar cars (e.g., foreign models or antique cars) are problematic, as this can mean moving outside the trained perceptual space. For example, expertise with modern cars, and the associated holistic processing style adopted, does not transfer to antique cars (Philips, Grovola, Bukach, & Gauthier, 2007). Similarly, faces from an unfamiliar race are not as well recognized and are not processed as holistically as own-race faces (Tanaka, Kiefer, & Bukach, 2004). However, previous studies that have not manipulated race have shown an inversion effect with both familiar and unfamiliar faces (Scapinelli & Yarmey, 1970; Yarmey, 1971). This suggests that holistic processing does not depend on familiarity with an exemplar, although it may be necessary for objects to come from an area of perceptual space that is very familiar to the observer.

From a practical standpoint, this makes the manipulation of familiarity more difficult with cars than with faces. That is, the car category is limited by the finite number of cars models that have ever been manufactured, and a very experienced car expert might easily notice when presented with a model that does not exist, even though it would fit in principle within the familiar perceptual space. In contrast, face experts cannot aspire to having experienced all possible faces, so there is nothing strange or particular about unfamiliar faces. In Experiment 4, we manipulated the familiarity of faces to see if we could further rule out contributions of pre-existing LTM representations to the VSTM advantage observed in Experiments 1-3. Notably, the nature of the VSTM advantage for faces and that for cars among car experts is remarkably similar not only in size, but also in its orientation-specificity and encoding time course. Therefore, the manipulation of face familiarity should have implications for effects of familiarity on VSTM more generally.

Familiar objects, whether they are faces of famous individuals or Toyota's best-selling sedan model, are typically distinguished both by the frequency with which they are seen and by the labels or semantic information associated with them. Thus, if such information can facilitate VSTM, we might expect better VSTM for famous faces – not because they are processed more holistically, but because of a contribution from semantically-related information. Although it would be reasonable to expect main effects of familiarity and of inversion (because of reduced holistic processing) on VSTM capacity, we predicted that there would not be an interaction between them. Such an interaction would be required to account for the orientation-specific VSTM advantage for cars among car experts, which did not emerge among novices. That is, if familiarity is a possible basis of the orientation-specific VSTM advantage for cars among car experts, it should increase VSTM for upright stimuli more than it does for inverted stimuli. In sum, Experiment 4 will allow us to test two important questions: does familiarity facilitate VSTM performance, and if so, can it account for the orientation-specific VSTM advantage for objects of expertise?

Methods

Participants

Thirty-one (4 male, age M = 19.12, SD = 2.02) participants from Temple University participated for course credit. All participants had normal or corrected-to-normal vision. Data from two participants were excluded prior to analysis due to poor performance (i.e. the VSTM capacity estimate, K, was equal to zero in at least one condition).

Stimuli

A total of 120 images were used, consisting of three front-on images of each of 40 famous individuals (actors or entertainers; 20 female, 20 male). All images had a neutral facial expression. To maximize the familiarity of the faces, only individuals who frequently appear in the current popular media (e.g., Tom Cruise, Tom Hanks, Julia Roberts, etc.) were chosen. Each image was cropped to remove hair and background, was converted to grayscale, and was scaled to equate image height. Each images was divided into three segments separating the eye, nose, and mouth regions of each face. Forty new “familiar” images of these famous individuals were created by combining a mouth, nose, and eye segment from each of the three images of the same famous person (Figure 5 A, left column). The parts were aligned in such a way that the configural relations between the different regions of the face were as identical as possible to those in the original images of the famous individuals. Forty additional “unfamiliar” face images were created by re-combining these same face segments, but in such a way that the different pieces within any one face came from three different famous individuals (Figure 5 A, right column). This manipulation of familiarity was necessary to control the content of the images across the familiar and unfamiliar conditions and thus eliminate spurious lower-level perceptual differences that could potentially impact VSTM. Therefore, what defined familiarity was the conjunction of the different parts of faces. To confirm the validity of our stimulus manipulation, twenty-one of the participants completed a survey after completing the VSTM task, in which they were asked to rate the familiarity of each face on a scale of 0 (completely unfamiliar) to 10 (highest possible familiarity). The mean familiarity ratings for the familiar (5.60) and unfamiliar (2.46) faces were significantly different, t(20)=7.99, p≤.0001.

Figure 5.

(A) Examples of the familiar (left column) and unfamiliar (right column) faces used in Experiment 4. Note that the unfamiliar faces in the right column share a feature with the familiar faces in the left column (that is, that of Elijah Wood and Julia Roberts). (B) Visual short-term memory (VSTM) performance in Experiment 4. VSTM was greater for familiar compared to unfamiliar faces regardless of orientation, with both types of stimuli experiencing a similar drop in performance with inversion.

Design, Procedure and Analysis

The design and procedure were similar to those used in Experiments 1& 2, except that the set size was fixed at 5 and the VSTM study array was always presented for 4000-ms. Participants performed 10 blocks of 16 trials. Forty trials were presented in each of the four conditions (i.e., upright famous faces, inverted famous faces, upright unfamiliar faces faces, inverted unfamiliar faces). Trials for each orientation were blocked and participants performed alternating blocks containing only upright or only inverted stimuli. Twenty famous faces (10 upright, 10 inverted) and 20 unfamiliar faces (10 upright, 10 inverted) appeared throughout the study. Stimuli were selected so that each of the face segments appeared in only one condition (i.e., either in the context of an upright famous face, inverted famous face, upright unfamiliar face, or an inverted unfamiliar face). In addition, in the unfamiliar face condition, where the different regions from the same face were separated across different face images, all the segments from the same original identity always appeared within the same condition (e.g. if Nicole Kidman's eyes appeared in the inverted condition for a participant, her nose and mouth also appeared inverted but across different unfamiliar face images). The allocation of the stimuli to each of these conditions was counterbalanced across participants.

Results

A 2 (orientation; upright, inverted) × 2 (familiarity; familiar, unfamiliar) ANOVA revealed main effects of orientation, F(1,28)=27.97, p≤.0001, and familiarity, F(1,28)=6.63, p=.016, but no interaction between these two variables, F(1,28)=0.63, p=.433. In sum, despite a general increase in VSTM for familiar faces over unfamiliar faces, the VSTM advantage for upright over inverted faces was similar for both familiar and unfamiliar faces (Figure 5 B).

Discussion

VSTM performance was impacted by the familiarity of face stimuli, suggesting that familiarity could also play a role in the larger VSTM capacity for cars among car experts compared to car novices. However, arguing against a familiarity account of the VSTM advantages for objects of expertise is the generalization of this familiarity advantage to both upright and inverted stimuli. Interestingly, in hindsight, a similar small familiarity advantage for cars among car experts (relative to car novices), irrespective of orientation, appears to have been present as a trend in the other studies reported earlier in this paper (e.g., p = .172 in Experiment 1, p = 0.11 in Experiment 2). Thus, although familiarity may play a general role in increasing VSTM capacity for both upright and inverted cars among car experts, it cannot account for the orientation-specific VSTM advantage for objects of expertise. Therefore, this study suggests that familiarity on its own cannot account for the expert advantage reported in the Experiments 1 – 3.

Previous studies exploring the effect of familiarity on the face inversion effect report results similar to that of ours. For example, Scapinello & Yarmey (1970) also report a lack of an interaction between inversion and familiarity in their study of face recognition. This study, however, used faces that became familiar through training within the lab. To explore whether these findings extend to familiarity gained in the real world, a follow-up study replicated the finding using faces of famous individuals in the familiar condition (Yarmey, 1971). The follow-up results were remarkably similar to those reported in Scapinello & Yarmey (1970), and although the author suggested that verbal labels typically associated with familiar faces may increase recognition performance more for upright than inverted faces, no such interaction between familiarity and inversion was present in the data.

The lack of an impact of familiarity on the orientation-specific nature of the VSTM advantage for objects of expertise is also consistent with findings from electrophysiological studies. The N170 electrophysiological potential has been robustly linked with the structural (holistic) encoding of faces and other objects of expertise (Gauthier et al., 2003; Rossion et al., 2000). Critically, this potential is modulated by inversion, but not familiarity (Bentin, 1996; Benton & Deouell, 2000; Eimer, 2000; Jemel, Schuller, Cheref-Kahn, Goffaux, Crommelinck, & Bruyer, 2003; Rossion et al., 2000; Rossion et al., 1999; but see Caharel et al., 2002, Marzi & Viggiano, 2007). Thus, the failure of our familiarity manipulation to impact the orientation-specific VSTM advantage for cars among car experts is consistent with a holistic processing locus for this advantage. The electrophysiological literature also provides evidence regarding the potential locus of the small VSTM advantage for familiar over unfamiliar faces irrespective of orientation: a later electrophysiological potential, the P250, is modulated by familiarity, showing greater amplitude for familiar than unfamiliar faces (Tanaka, Curran, Porterfield, & Collins, 2006; Pfutze, Sommer, & Schweinberger, 2002; Schweinberger, Pickering, & Jentzsch, 2002; Begleiter, Porjesz, & Wang, 1995; Schweinberger, Pfutze, & Sommer, 1995). Notably, it has been suggested that this potential is the earliest component associated with a stored perceptual representation in long-term memory (Pfutze, et al., 2002; Schweinberger et al., 2002). Intriguingly, a recent finding suggests that while this later P250 component is modulated by familiarity, it is less sensitive to inversion (Marzi & Viggiano, 2007). Thus, further studies might look to the P250 for additional insight into the impact of familiarity on VSTM.

General Discussion

The VSTM advantage for cars among car experts is remarkably similar to that demonstrated for faces; this advantage requires sufficient encoding time, is orientation-specific and is similar in magnitude to the VSTM advantage for faces. These findings are consistent with a general perceptual expertise account of the VSTM advantage for faces. This advantage was not eliminated by the introduction of a verbal memory load previously demonstrated to impact verbal short-term memory performance, suggesting that it does not rely on verbal short-term memory. Nor was it eliminated by the use of a small stimulus set, which increased the potential for proactive interference on LTM recall, or by a surprise switch in stimulus set, which probed for any advantages due to stimulus-specific representations in LTM. Finally, a role of real-world familiarity, or more specifically the resulting LTM representations, in producing the orientation-specific VSTM advantage for objects of expertise was ruled out despite evidence that familiarity could produce a general boost to VSTM performance irrespective of object orientation. This expert advantage is robust and does not seem to depend on the direct recruitment of additional capacity from other memory systems such as verbal or long-term memory.

We suggest that the mechanism underlying this expert VSTM advantage likely involves holistic processing, which is common to the processing of faces and of cars among car experts. The correlation between VSTM for cars and sensitivity on an established measure of car expertise is consistent with such an account: this car expertise index is correlated with measures of holistic processing of cars, and it is correlated with the N170 electrophysiological potential (Gauthier et al., 2003), which is modulated by inversion but not familiarity. The orientation-specific nature of this advantage is also consistent with a contribution from holistic processing mechanisms: the inversion effect for faces is thought to result from reduced access to configural information critical for holistic processing (Collishaw & Hole, 2002; Kemp, McManus, & Pigott, 1990; Leder & Bruce, 2000; Leder, Candrian, Huber, & Bruce, 2001; Murray, Yong, & Rhodes, 2000; Rhodes, Brake, & Atkinson, 1993; Searcy & Bartlett, 1996; Tanaka & Sengco, 1997; Thompson, 1980). Thus, many aspects of the VSTM advantage for objects from one's domain of expertise suggest that it may be driven by holistic encoding strategies recruited by visual experts, as in the case of expert face perception.

The robustness of holistic processing effects for non-face objects of expertise has been questioned recently (McKone, Kanwisher, & Duchaine, 2006), so the current study also provided an important test of the relationship between inversion effects and expertise. Contrary to suggestions that the inversion effect is specific to face processing, the results of Experiments 1 and 2 show robust inversion effects for cars in car experts of the same magnitude as that found with faces⁸. This was not the case among novices. Thus, the orientation specificity of car experts' VSTM advantage for cars is consistent with a holistic processing account of this advantage as well as of perceptual expertise more generally.

Holistic encoding may benefit VSTM capacity by providing a tighter binding of information in object representations. Feature-based theories of VSTM suggest that VSTM is limited by both the capacity of independent feature stores and the capacity of attentional mechanisms required to maintain the binding between features (Delvenne & Bruyer, 2004; Wheeler & Treisman, 2002). Holistic processing may integrate spatially separate features into the same feature unit, thereby using fewer feature slots and reducing the burden on attentional resources. Similarly, consistent with object-based theories of VSTM capacity, holistic encoding may allow experts to incorporate more features into the unified object representations suggested to serve as the units of VSTM. This could be especially beneficial for discriminating highly similar exemplars of complex objects, such as cars. Therefore, holistic processing may allow experts to maximize the use of an inherently limited VSTM system.

At a finer scale, VSTM is influenced by the organization of features within objects, providing another avenue for holistic encoding to impact VSTM performance. VSTM for features is improved when the features come from the same part, rather than different parts, of an object (Xu, 2002). For example, color and orientation information are best encoded when they are from the same part of an object, less well encoded when they are from different parts of the same object, and least well encoded when they are from spatially separated objects (Xu, 2002). In the context of the holistic processing strategy recruited by visual experts for objects within their domain of expertise, features from what would be considered different parts of an object by novices may be encoded and represented as being from the same part by experts. Thus, this would allow experts to take advantage of the part benefit for feature encoding (Xu, 2002). More specifically, the more integrated nature of representations of faces, and cars among car experts, may underlie the VTSM advantage demonstrated for objects of expertise. Therefore, the influence of object-based hierarchical feature encoding on VSTM capacity provides another avenue whereby differences in the perceived relationship between features in holistic versus feature-based object representations may impact VSTM capacity. Future work will need to test these hypotheses more directly.

Recent neuroimaging studies provide further insight into the system underlying VSTM and thus possible loci for the expertise effect on VSTM capacity (Song & Jiang, 2006; Xu & Chun, 2006). Such studies not only provide evidence of dissociable roles for the different nodes in the system supporting VSTM for objects, but they also report a neural correlate of the effect of complexity on VSTM capacity. A number of core areas spanning frontal, parietal, occipital and temporal lobes have been implicated in VSTM (Desimone, 1996; Druzgal & D'Esposito, 2001; Druzgal & D'Esposito, 2003; Pessoa, Gutierrez, Bandettini, & Ungerleider, 2002; Todd & Marois, 2004). Xu and Chun (2006) found that while activity patterns in the inferior intraparietal sulcus (IPS) suggest that this node of the VSTM system has a fixed capacity of about four objects (regardless of object complexity), activity patterns in the superior IPS and the lateral occipital complex (LOC) suggest that the capacity of these areas is variable, depending on the complexity of the objects stored. Thus, it appears that a complexity induced bottleneck in the superior IPS and LOC leads to the observed lower VSTM capacity for complex objects. Xu and Chun (2006) suggest that the inferior intraparietal sulcus is responsible for maintaining spatial attention over a fixed number of objects, while the superior parietal sulcus and the lateral occipital complex are important for encoding and maintaining the specific object representations.

The involvement of object form processing areas in the effect of complexity on VSTM capacity (Song and Jiang, 2006; Xu and Chun, 2006) is consistent with the suggestion that the locus of the expert VSTM effect may be in the nearby fusiform region, namely the FFA, implicated in the perceptual processing of objects of expertise (Gauthier et al., 2005; Gauthier et al., 2000; Gauthier, Tarr, Anderson, Skudlarski, & Gore, 1999; Xu, 2005). Importantly, previous studies have already implicated the FFA in VSTM for faces (Druzgal & D'Esposito, 2001; Druzgal & D'Esposito, 2003). In addition, the level of activity in the FFA after expertise training with a novel category of objects is correlated with behavioral measures of holistic processing (Gauthier et. al., 1999). Thus, the recruitment of the FFA for objects of expertise and the resulting holistic processing strategy may allow experts to better encode complex visual information, potentially reducing the perceived complexity of objects of expertise. This may limit the susceptibility of expert VSTM capacity to object complexity. If true, this would suggest that activation in the superior IPS and/or occipital/temporal areas responsible for encoding information in VSTM reflects the perceived, rather than physical, complexity of the objects stored in VSTM. Future studies will explore this prediction.

The findings reported in this study also provide an alternative to the recent claim that the ‘true’ capacity of VSTM, free from contamination by LTM, verbal memory, or contexual information is limited to one object (Olsson & Poom, 2005). Olsson & Poom (2005) found that with 500-ms of encoding time, participants had a VSTM capacity for intra-categorical geometric shapes (e.g., ovals with varying aspect ratios) of only a single item. Based on this finding, they suggest that performance in previous studies reporting a VSTM capacity of 3-4 objects (e.g. Luck & Vogel, 1997; Vogel et al., 2001) was facilitated by categorical structures in LTM. Specifically, they suggest that such a benefit arises from the use of stimuli that cross category boundaries (e.g. a red and a yellow square cross a color boundary). In the studies reported here, the faces were unfamiliar, with no obvious labels and belonged to a single category. Therefore, according to Olsson and Poom (2005), observers should have had a capacity of only a single face under such conditions. One possible reason for this inconsistency is the limited encoding time in the Olsson and Poom (2005) study; our findings and those of Eng et al. (2005) and Curby & Gauthier (2007) suggest that VSTM capacity for complex objects is underestimated with 500-ms of encoding time because of perceptual encoding limitations. It is possible that capacity for the geometrical objects used by Olsson & Poom (2005) could reach that reported for cars among novices, for instance, given enough encoding time.

In conclusion, we provide evidence that VSTM capacity for complex objects is not hard-wired, and instead can be influenced by one's experience—or more specifically, we suggest that the nature of representations stored in VSTM allows visual experts to maximize the storage efficiency of an otherwise inherently limited system. Thus, extensive experience with a category of objects, such as that required to produce the qualitative shift in encoding strategy seen among perceptual experts, leads to greater VSTM than would be expected based on the complexity of the objects stored. It remains an empirical question as to whether other types of encoding strategies, besides holistic, benefit VSTM capacity.