Spatial scaling factors explain eccentricity effects on face ERPs

Abstract

Event-related potential (ERP) studies consistently have described a strong, face-sensitive response termed the N170. This component is maximal at the fovea and decreases strongly with eccentricity, a result that could suggest a foveal bias in the cortical generators responsible for face processing. Here we demonstrate that scaling stimulus size according to V1 cortical magnification factor can virtually eliminate face-related eccentricity effects, indicating that eccentricity effects on face ERPs are largely due to low-level visual factors rather than high-level cortical specialization for foveal stimuli.

Introduction

The degradation of visual performance with eccentricity strongly limits our capacity to apprehend objects such as faces in natural scenes (Rousselet, Thorpe, & Fabre-Thorpe, 2004). Eccentricity effects might be the direct consequence of low-level factors (Banks, Sekuler, & Anderson, 1991; Bennett & Banks, 1991; Mäkelä, Näsänen, Rovamo, & Melmoth, 2001). Alternatively, they could reflect the specialization of object-selective cortical areas for foveal stimuli (Levy, Hasson, Avidan, et al., 2001; Hasson, Levy, Behrmann, Hendler, & Malach, 2002). In keeping with this last idea, the early selective response of the visual cortex to faces (time range 140–200 ms), as indexed by central (VPP: Jeffreys, 1996) and posterior (N170: Carmel & Bentin, 2002; Itier & Taylor, 2004a; Rossion et al., 2000; Rousselet, Macé, & Fabre-Thorpe, 2004) event-related potentials (ERP), has been shown to decline dramatically when stimuli are presented few degrees away from fixation (Eimer, 2000; Jeffreys, Tukmachi, & Rockley, 1992). Such a decline is consistent with the notion of selective tuning of face processing areas to foveally presented faces. However, before we can accept the theory of a foveal bias in face and object processing, we first must rule out the possibility that eccentricity-based effects may be a simple consequence of the reduced cortical representation of peripherally presented stimuli (Mäkelä et al., 2001). One obvious manipulation to control for reduced cortical representation is to magnify stimulus size according to the cortical magnification factor. 

Jeffreys et al. (1992) manipulated stimulus size and found that an eightfold size increase did influence the degree to which the ERP response to faces (VPP) was attenuated by eccentricity. However, even using such large faces (9° in height) the VPP was very weak, if present at all, when faces were centered at 4° of eccentricity and more (Jeffreys, 1996; Jeffreys et al., 1992). This result led Jeffreys to conclude that the VPP was an automatic response to fixated faces. Eimer (2000) studied the effect of eccentricity on the N170 and the VPP. Compared to central presentations, Eimer found a reduced but reliable N170 difference between faces and houses centered at 3.5° from the fixation point. The VPP, on the other hand, was no longer present at 3.5°. Eimer did not manipulate stimulus size. 

In the present experiment, we examined the effect of eccentricity on the N170, an ERP component very sensitive to faces, controlling for the effect of cortical magnification. Although there has been an extensive debate in the literature as to whether the VPP and the N170 reflect the activity of the same generators (e.g., Eimer, 2000), there is now strong evidence that they reflect the opposite sides of the same coin (e.g., George, Evans, Fiori, Davidoff, & Renault, 1996; Itier & Taylor, 2002, 2004a; Jemel et al., 2003; see particularly Joyce & Rossion, 2005). Fifteen subjects viewed a series of randomly interleaved faces and houses presented either centrally or centered at 5° or 10° to the left or right of fixation (abbreviations: 0°, L5°, R5°, L10°, R10°). Stimuli were matched for spatial frequency content by averaging their amplitude spectrum. This manipulation was introduced to rule out the hypothesis that the face eccentricity effect might be due to low-level differences in the spatial frequency content of houses and faces. Peripheral stimuli were presented at one of two sizes: either matched to the central presentation size, or scaled to compensate for differences in V1 cortical representation (Figure 1). Our results show that scaling stimulus size according to the V1 cortical magnification factor can virtually eliminate face-related eccentricity effects. This demonstrates that eccentricity effects on face ERPs are largely due to low-level visual factors rather than to high-level cortical specialization for foveal stimuli. 

Figure 1

 

Examples of two stimuli at different sizes used in the experiment. The size ratio has been respected in this figure.

Methods

Subjects

Fifteen subjects participated in this experiment. All subjects gave written informed consent and had normal or corrected-to-normal vision. Subjects ranged in age from 19 to 32 (mean age 25), and received $10/hour for their participation. Twelve subjects were right handed, and seven were female. 

Stimuli

Stimulus construction

Faces were front-view photographs converted to greyscale, and cropped within a common oval frame that removed both hair and face contour information. Each of the 10 cropped faces was 3.0° high and 1.8° wide in the nonscaled stimulus, and the face was centered in a 3.0° × 3.0° background of average luminance (for more details, see Gold, Bennett, & Sekuler, 1999). The 10 house stimuli were derived from front-view photographs of houses converted to greyscale. The house set employed in the experiment was designed to share the configural homogeneity typical of faces (i.e., with key features found in the same relative configuration across exemplars). To replicate this homogeneity within the house set, one house was chosen to act as a base house for all stimuli. This house was cropped from the background scene information (bushes, driveway, etc.) and placed on a light grey background. The base house contained a door in the lower left, a large window in the lower right, and a pair of upper windows. To create the individual houses in the house set, the windows and doors were replaced with new exemplars from other photographs, placed in the same position as the originals. As with faces, each individual house could be discriminated based on any one of these individual features. The faces and houses were then placed on backgrounds of uniform grey (luminance = 20.2 cd/m²). These final stimuli were then equated in terms of spatial frequency content by taking the average of the amplitude spectra of all 20 stimuli and then combining that average spectrum and the original phase spectra to reconstruct each individual stimulus. Because form information is largely carried by phase rather than amplitude (Oppenheim & Lim, 1981; Sekuler & Bennett, 1996), individual houses and faces remain easily discriminable after this manipulation. However, this manipulation ensures that any eccentricity-related differences in the EEG to faces or houses are not simply a function of differences in the relative visibility of specific frequency components in the stimuli. 

Determining stimulus size

To determine the amount an image should be magnified on the screen so that it stimulates a constant cortical area, the ratio of cortical magnification (M) at fixation versus magnification at eccentricity E needs to be determined. To do so we used the following formula (Horton & Hoyt, 1991): M_linear=A/(E+e2), with E the eccentricity in degrees, A the cortical scaling factor in mm, and e2 the eccentricity in degrees at which a stimulus subtends half the cortical distance that it subtends at the fovea. We used A=29.2 mm and e2=3.67° based on a recent report of the cortical magnification factor in V1 (Dougherty et al., 2003). For a stimulus presented at fixation, E=0° and M=29.2/3.67=7.96. For a stimulus presented at 5° from fixation, E=5 and M=29.2/(5+3.67)=3.37. So, the image at fixation stimulates a cortical surface area that is 7.96/3.37=2.36 times larger than the surface stimulated by the same image when it is presented at 5°. Therefore, the image size had to be multiplied by 2.36 to compensate for the magnification factor when the image was presented at 5°. Finally, for a stimulus presented at 10° from fixation, E=10 and M=29.2/(10+3.67)=2.14. By the same reasoning, image size had to be multiplied by 7.96/2.14=3.72 when the image was presented at 10°. Based on screen constraints, we choose the largest face height to be 350 pixels (11.2°, the width of the cropped stimuli being 6.8°). The size for a central stimulus was thus 350/3.72=94 pixels (3.0°, the width of the cropped stimuli being 1.8°). At 5° it was 94×2.36=222 pixels (7.2°, the width of the cropped stimuli being 4.4°). There were a total of 18 conditions in this experiment: faces and houses were seen at five different positions, and at four of these positions they could have two different sizes. 

Experimental design

Subjects sat in a dimly lit sound-attenuated booth. Viewing distance was maintained at 70 cm by the use of a chin rest. Stimuli were presented for 80 ms (six frames at 75 Hz) on a Sony Trinitron GDM-F520 monitor (1024 × 768 pixels, effective height and width: 40.5 × 30.5 cm). Subjects had to respond by pressing one of two keys to indicate whether a face or house appeared on the screen. The button/category association was counterbalanced across subjects. An experiment was composed of two sessions, each containing 1080 trials (90 trials × 12 blocks). Each session used 5 of 10 exemplars from each object category (order was counterbalanced across subjects). Within each block, there was only one presentation of each item (face or house) in each of the nine conditions (0°, L5°, L10°, R5°, R10°, magL5°, magL10°, magR5°, and magR10°). A trial was organized as follows: A blank screen was presented for about 200 ms, followed by a small fixation cross (a 0.3° ‘+’ in the middle of the screen) for a random duration ranging from 500 to 900 ms). Then a stimulus was presented for 80 ms, followed by a blank screen for 1000 ms during which time subjects were allowed to make a response to the categorization task (face or house). After that delay, responses were considered incorrect. Trial durations thus ranged from 1780 to 2180 ms. 

EEG data acquisition and processing

EEG data were acquired with a 256-channel Geodesic Sensor Net (Electrical Geodesics Inc., Eugene, Oregon; Tucker, 1993). Analog signal was amplified about 1000 times, digitized at 500 Hz, and band-pass filtered between 0.1 and 200 Hz. The ground electrode was placed along the midline, ahead of Fz, and impedances were kept below 50 kΩ. Subjects were asked to minimize blinking, head movement, and swallowing. Subjects were then given a description of the task. They were carefully instructed about the importance of maintaining fixation even when stimuli were peripherally presented. Electrode positions were obtained by means of a Polhemus spatial digitizer. 

The main experiment was preceded and followed by a short eye calibration procedure that was aimed at obtaining ERP topographies for prototypical eye movements. These topographies were used during data analysis to correct for eye movement artifacts (see BESA artifact correction manual). In randomly interleaved blocks, subjects were asked to move their eyes left, right, up, and down, and to blink. Each block was composed of 20 trials, providing a total of 40 trials per condition across the two sessions. In all conditions, a white point (0.1°, 112 cd/m²) was presented against a black background. For eye movements, a white point was alternatively presented in the middle of the screen for 1000 ms and then 10° away from the fixation point for 1000 ms. Subjects were instructed to always maintain fixation on the point, to follow the point with their eyes when it moved, but to avoid anticipatory movements, and blinking during eye movements. In a blink block, the point was presented for 1000 ms followed by a blank screen for 1000 ms. Subjects were instructed to blink once whenever the point flashed off. 

EEG analysis was performed using BESA 5.0 (MEGIS software GmbH). EEG data were referenced on-line to electrode Cz and re-referenced off-line by subtracting the average of all signals from each individual signal. The signal was then band-pass filtered in the range 1–30 Hz and bad channels interpolated. Baseline correction was performed using the 200-ms of prestimulus activity. Two artifact rejections were applied on all electrodes over the [−200 ms; +300 ms] time period, first with a criterion of [−100; +100 μV] to reject trials with excessive amplitude, second to reject trials in which the difference between two consecutive time points was greater than 75 μV. Only correct trials were averaged. ERPs were corrected for blink and horizontal movement artifacts before further analysis. 

Source modeling in BESA

The actual location of the N170 sources is still controversial and seems to correspond to the activity of a distributed network including at least medial temporal areas and ventral occipital–temporal areas (Bötzel, Schulze, & Stodieck, 1995; Itier & Taylor, 2004b; Watanabe, Kakigi, & Puce, 2003). However, the goal of our modeling was not to provide an accurate estimate of the sources of the N170, but rather to provide a compact description of the signal recorded over the entire scalp. We used separate averages of the uncorrected ERP and of the average artifacts to create independent spatial components for the ERP and the artifacts (see BESA artifact correction manual for details). The N170 was modeled using two symmetric regional sources over the interval 140–240 ms after stimulus onset. The positions of the sources generally were compatible with generators situated in extra-striate visual cortex, but there was significant variability across subjects. Each source was composed of three orthogonal vectors that were squared and added together before measuring the mean amplitude of the source waveforms over the time interval 140–240 ms. We performed analyses on the residual variance of the model for the different conditions. For both scaled and nonscaled stimuli, an ANOVA revealed a main position effect (both p < .005), but no category effect and no interaction between category and position factors. For nonscaled stimuli, post hoc paired t tests revealed significant differences between central presentations and presentations at 10° (left and right) and 5° right. For scaled stimuli, there were only significant differences between central presentation and presentations at 10° (left and right). Those effects were expected given the drop of signal to noise ratio with increased eccentricity. Statistical analysis on the 3D coordinates of the regional sources in the face and house conditions did not reveal any significant effect. 

Results

At the behavioral level, subjects performed very well in this task with a mean accuracy of 92.2% across conditions (range 88.4–94.7%). For reaction times (mean 517 ms), there was an interaction between eccentricity and scaling (Figure 2), with no effect of eccentricity for scaled stimuli, but slower responses for peripherally presented nonscaled stimuli, F(2.0,28.2) = 26.3, p < .0001. 

Figure 2

 

Mean reaction times as a function of eccentricity and size.

As found previously (Eimer, 1998, 2000; Itier & Taylor, 2004a; Rossion et al. 2000), the foveal N170 was larger in amplitude for faces than for houses and was characterized by a central-positive and bilateral posterior-negative topography (Figure 3). In keeping with the broader literature, we refer to the amplitude difference between faces and houses (“the face effect”) as a marker for face processing. Notably, this is the first time the face effect has been shown for stimuli equated in spatial frequency content. Furthermore, contrary to Eimer (2000) and Jeffreys et al. (1992; Jeffreys, 1996) who found no VPP for faces centered respectively at 3.5° and 4°, our data clearly show both an N170 and a VPP at all eccentricities (Figure 3). This constitutes another argument in favor of a common origin of these two components (Joyce & Rossion, 2005). However, a complete discussion of the potential relationship between the N170 and the VPP is beyond the scope of this paper. The reason that Jeffreys did not record a VPP even for large faces presented 4° or more from the fixation point might be that the very few subjects he tested presented a pattern we observed in some of our subjects, wherein the VPP topography changes slightly with eccentricity, presenting a maximum over frontal and central electrodes ipsilateral to the stimulation. Because Jeffreys recorded only from central electrodes, he could have missed the lateralized VPP. 

Figure 3

 

Interpolated 2D topographical maps of the differential activity between face and house ERPs. Each map represents the average activity over a time window indicated at the bottom of the figure. Nose is pointing upward. For each map, a cross is centered at electrode Cz (situated at the apex of the scalp). The circle joins the glabella (brow ridge) and the occipital protuberance (back of the skull). The two holes on the left and right sides of the maps correspond to the locations of the ears. In the left columns, all stimuli were the same size. In the right columns, stimuli were cortically magnified, except central stimuli (C) whose maps are identical to those presented in the left columns. L5, R5, L10, and R10 refer to stimuli presented in the left/right visual field at 5/10° of eccentricity. Interpolated ERP maps were made with the CarTool data analysis software (3.1) developed by Denis Brunet at the Functional Brain Mapping Laboratory, Geneva, Switzerland.

The time course of the face effect was determined at each eccentricity by comparing the global field power (GFP) for faces and houses at each time point (Figure 4). GFP is a measure of changes in electric field strength and is computed as the spatial standard deviation of the scalp electric field (Lehmann & Skrandies, 1980). Stronger electric fields lead to larger GFP values. It is assumed that peaks of GFP coincide with maximum activation of the underlying generators. Significant GFP differences between two conditions were assessed at each time point and at each scalp location using bootstrap methods (999 permutations, p < .01, corrected for multiple comparisons; Figure 4). This analysis revealed a significant difference starting at 166 ms for central presentations. For nonscaled stimuli, the effect appeared slightly later in the L5° (184 ms) and R5° (170 ms) conditions and was no longer significant at 10°. However, there was a dramatic change in the results when faces and houses were enlarged to compensate for cortical magnification differences. First, the difference in GFP strength between faces and houses presented at 5° appeared at about the same latency that was observed at 0° (L5° = 156 ms; R5° = 164 ms). Second, this effect re-emerged at 10°, with a similar time course as that found at the other eccentricities (L10° = 160 ms; R10° = 158 ms). 

Figure 4

 

Time course of the global field power (GFP) for faces and houses in the different conditions. The GFP is the instantaneous spatial standard deviation computed across electrodes at each time point (Lehmann & Skrandies, 1980). This measure is independent of the reference electrode and the number of electrodes and thus provides a compact description of the signal across the head. Time points at which signals for faces and houses diverged significantly (p < .01, 999 bootstrap permutations) are indicated by red points along the horizontal axis.

The same pattern of results was obtained when comparisons were performed on the ERP at individual electrodes rather than on the GFP. Significant ERP differences between two conditions were assessed at each time point and at each scalp location using bootstrap methods (999 permutations, p < .01, corrected for multiple comparisons). Clear differences were observed in the 0°, L5°, and R5° nonscaled conditions at posterior, central, and frontal electrodes (Figures 5 and 6). Only very few of those electrodes presented significant differences at 10°, and those differences were very brief in time with onsets after 200 ms. Scaling stimulus size again virtually eliminated the eccentricity effect. 

Figure 5

 

Grand average ERPs averaged across a cluster of nine neighboring posterior electrodes at which the signal was maximal. Left electrodes = 83-84-85-94-95-96-104-105-106; right electrodes = 162-163-170-171-172-178-179-180-190. Face and house ERPs are depicted in red and blue traces respectively. Left (L) and right (R) hemisphere ERPs are depicted in dotted and solid lines, respectively. The N170 decreased dramatically in amplitude with eccentricity and progressively lost its face specificity, an effect that could be restored by scaling stimulus size. Note also the clear lateralization of the N170, first recorded at electrodes over the hemisphere contralateral to stimulus presentation.

Figure 6

 

Significant differences at each electrode over time between face and house ERPs. Differences were assessed at each time point and each electrode using 999 bootstrap permutations (p < .01, corrected for multiple comparisons). Significant time points are indicated in red (|face ERP| > |house ERP|) and blue (|house ERP| > |face ERP|) while nonsignificant time points are indicated in grey. Electrodes are stacked along the vertical axis. The horizontal black lines separate the different groups of electrodes organized in frontal, central, and posterior electrodes (F/C/P) and subdivided into left hemisphere, median, and right hemisphere electrodes (L/M/R).

To estimate the activity of the brain sources in the different conditions, without biasing the analysis toward specific electrodes, the ERP signal was modeled in BESA by using two regional sources constrained in symmetry over the time window 140–240 ms (Figure 7). The model accounted for 85.4–91.7% of the variance in each condition and the goodness-of-fit did not vary significantly between faces and houses. For nonscaled stimuli, there was a strong decrease in the amplitude of the modeled sources with eccentricity for both faces and houses, and a decrease in the differential amplitude between faces and houses. Indeed, whereas there was a strong difference between faces and houses at the fovea, this difference was already reduced at 5° and not significant at 10°. For scaled stimuli, differences between faces and houses were similar at all positions, indicating that scaling stimulus size eliminated the eccentricity effect. 

Figure 7

 

Mean activity of the regional sources as a function of eccentricity. Measurements were made on the power of the source waveforms in the time window 140–240 ms after stimulus onset and entered into ANOVAs with category (2), position (5), and hemisphere (2) as within-subject factors. Nonscaled stimuli (dashed lines): position effect, F(2,28) = 10.2, p < .0001; Position × Category interaction, F(3,43) = 9.8, p < .0001. Scaled stimuli (solid lines): main category effect, F(1,14) = 27.7, p < .0001; no interaction. These analyses did not reveal any hemispheric lateralization effect.

Conclusions

The present ERP results rule out the hypothesis that the face eccentricity effect is due to low-level differences in the global spatial frequency content of houses and faces. Furthermore, we find no evidence that there is a foveal bias for face processing (as indexed by the N170) per se. Rather, eccentricity-based differences in face processing appear to be largely attributable to cortical magnification, in keeping with recent results from the monkey literature (Rolls, Aggelopoulos, & Zheng, 2003; Rousselet et al., 2004). However, even after magnification, the onset of the face effect was still slightly delayed for stimuli presented at 10° of eccentricity compared to the foveal condition (Figure 6; although note that the GFP results show, if anything, an earlier onset for scaled peripheral stimuli compared to central stimuli; Figure 4). Of course, the magnification factor we used was appropriate for V1, and so likely was not a precise match for the locus of the N170. Moreover, because other factors like contrast sensitivity and phase discrimination also change with eccentricity (Banks et al., 1991; Bennett & Banks, 1991; Mäkelä et al., 2001) and might affect face and object ERPs, a single spatial scaling factor is unlikely to account for all differences between object ERPs at the fovea and in the periphery. Also, stimulus scaling was not sufficient to compensate completely for the decrease of the absolute amplitude of the N170 and other ERP components. This result is difficult to interpret based on the monkey literature because, to our knowledge, no study comparable to the present experiment has yet been performed. But in keeping with the monkey literature, we believe it is more instructive to focus on differential activities, which might be a good indicator of the capacity of a large neuronal population (in the case of the ERP) to discriminate between two sets of stimuli. In addition, the difference in the P2 range was almost nonexistent at 10° for scaled stimuli (about 200–250 ms; Figures 4 and 6). This might indicate that late responding mechanisms, potentially related to face recognition rather than simple object discrimination (Halit, de Hann, & Johnson, 2000), are still perturbed after image scaling. However, any interpretation regarding the later component is necessarily tentative, as the current study examined discrimination rather than recognition. 

Overall, though, in the present experiment cortical magnification accounts for most of the eccentricity-related differences in the ERP response amplitudes to faces and houses. It is not yet clear whether these results contradict the model of Levy et al., which suggests that in fMRI the cortical areas responding more strongly to faces have a bias toward the central visual field while areas responding more strongly to houses have a bias toward the peripheral visual field (Hasson et al., 2002; Levy et al. 2001). Here we found a similar drop of ERP activity with eccentricity for both categories (Figure 7), whereas the model by Levy et al. would actually predict an increase of the response for houses in the periphery. However, the sources of the N170 are still debated (Itier & Taylor, 2004b; Watanabe et al., 2003) and fMRI has a relatively poor temporal resolution. Thus, the results by Levy et al. could reflect a later part of the neuronal response. In any case, our results strongly suggest that a high-level cortical specialization for foveal stimuli is not likely to be a rule applying to all face processing cortical areas. 

Finally, the present ERP results rule out the hypothesis that the face eccentricity effect is due to low-level differences in the global spatial frequency content of houses and faces. Because amplitude spectra were equated across all stimuli, the eccentricity-related differences in the EEG to unscaled faces and houses cannot simply be a function of differences in the relative visibility of specific frequency components in the two stimulus sets. As well, our results demonstrate that face and house processing is eccentricity and size dependent. This strong constraint needs to be taken into account by models of object processing in natural scenes, which typically ignore such low-level constraints (Rousselet et al., 2004). Our data are also in keeping with the idea that, at least in terms of face processing, the differences between the peripheral and the central visual fields are quantitative rather than qualitative in nature (Mäkelä et al., 2001). 

Acknowledgments

We acknowledge support from NSERC Discovery Grants 42133 and 105494, the CIHR fellowship program, and the Canada Research Chair program. McMaster University Research Ethics Board approved this work. 

Commercial relationships: none. 

Corresponding author: Guillaume A. Rousselet. 

Email: [e-mail]. 

Address: 1280 Main Street West, Hamilton, ON L8S4K1.