Direct neurophysiological evidence for spatial suppression surrounding the focus of attention in vision

Results

Exp. 1. While fixating the center of the screen, observers searched for a red target C among eight blue distractor Cs presented at a constant distance from fixation in the lower right quadrant (Fig. 1a). The target C appeared randomly at one of the nine possible stimulus locations (Fig. 1c), thus, forcing subjects to change the spatial focus of attention from trial to trial. On half of the trials, a white ring (the probe stimulus) was flashed around the center C 250 ms after the search frame onset (see Materials and Methods for details). In the other half of the trials, no probe was presented. Because the probe position was constant and the target position varied, there were five target-to-probe distances, ranging from probe distance 0 (PD0; target at probed location) through probe-distance 4 (PD4; target four items away from probe; see Fig. 1c).

To isolate event-related magnetic field (ERMF) response elicited by the probe from ERMF activity reflecting target processing, the ERMF response triggered by trials with a search frame but no probe stimulus (frame-only or FO trials) was subtracted from the ERMF response triggered by trials with a search frame followed by a probe (frame-plus-probe or FP trials). Previous studies have validated this approach (32–34). Fig. 2a shows the ERMF distributions of the resulting difference waves between 130–150 ms for each of the nine target locations, averaged across observers. The FP-minus-FO difference waveforms are shown for each target-to-probe distance in Fig. 2b (collapsed across mirror-symmetrical positions). Both the distributions and the waveforms show that the initial response was largest when the target location was probed (PD0, position 5), smallest when the location adjacent to the target was probed (PD1, positions 4 and 6), and intermediate for larger target-to-probe distances (PD2–PD4, positions 1–3 and 7–9). Thus, we observed an enhancement at the target location and a suppression at the adjacent location compared with the more distant locations, demonstrating that a narrow zone of attenuated cortical responsiveness surrounds the attended item. When inspecting the waveforms in Fig. 2b, it seems that there is also a small delay in latency for PD1 relative to PD0 and the other probe-distance conditions. However, a statistical comparison of the peak-latency for the PD0 and the PD1 conditions indicated that this difference was not statistically significant (F_1,11 = 2.5, P = 0.14).

The bar graph in Fig. 2c summarizes the magnitude of the average ERMF effect, using the difference between each observer's individual efflux and influx maxima to measure the size of the effect. Again, attending to the location adjacent to the probe's location led to a smaller probe response than attending to the probe's location or attending one location farther from the probe. For statistical validation of this activity profile, the magnitude of the magnetic field from 130–150 ms was subjected to an overall one-way repeated measures ANOVA (RANOVA) with a factor of target-to-probe distance (five levels: PD0–PD4). This analysis revealed a significant effect of target-to-probe distance (F_4,8 = 4.08, P < 0.05). Planned follow-up RANOVAs showed that the probe response was significantly smaller when the probe was at the location adjacent to the target location than when it was at the target location or two locations away from the target (PD1 vs. PD0: F_1,11 = 8.9, P < 0.05; PD1 vs. PD2: F_1,11 = 14.4, P < 0.01). This pattern demonstrates both inhibition at the location adjacent to the target and a recovery at more distant locations.

To evaluate the observers' behavioral performance, response time (RT) and percent error were subjected to two-way RANO-VAs with factors of target location (PD0–PD4) and probe presence/absence. The observers' responses to the targets were faster for no-probe trials (mean, 539 ms) than for probe-trials (mean, 543 ms), as reflected by a significant main effect of probe presence/absence (F_1,11 = 7.0, P < 0.05); this result presumably reflects backward masking of the target by the probe. There was also a significant main effect of target position on RT (F_4,8 = 6.0, P < 0.05) and a marginally significant interaction between target position and probe presence/absence (P = 0.08). Subsequent post hoc pairwise comparisons between the target positions revealed significant RT effects (P < 0.05) when comparing PD0 versus each of the other target locations (PD1–PD4). For percent error, no significant effects were observed.

The Cortical Locus of Attentional Attenuation. In this section, we describe the neuroanatomical origins of the surround suppression. However, it should be noted that the present data do not make it possible to completely isolate the suppressive effect from excitatory effects of attention. This limitation arises because the attenuation can be measured only as a difference between the response at PD1 and the response at other target-to-probe distances, and some excitatory effect may be present at these other distances. However, given the very large size of the suppressive effect, a reasonably precise measure of the inhibitory effect can be obtained by subtracting the FP-minus-FO response at PD4 from the FP-minus-FO response at PD1. Similarly, a reasonably precise measure of the excitatory effect can be obtained by comparing the PD4 and PD0 positions.

Fig. 3 shows the current source density estimates (SDEs) for the excitatory effect (Fig. 3a) and the inhibitory effect (Fig. 3b), averaged between 130 and 150 ms. Although the excitatory effect seems larger than the inhibitory effect at the chosen threshold, both effects were broadly distributed over the occipital lobe with a somewhat greater magnitude over the left (contralateral) hemisphere. This sort of broad distribution is consistent with previous studies showing attention effects across a wide swath of visual cortex (7, 8, 10, 35). The present data do not have the resolution to provide a more detailed localization, but it is clear that the attention-based surround suppression is present in visual cortex. Thus, this study provides direct evidence that the attended object is surrounded by a suppressive zone in early-to-intermediate areas of visual cortex that arises within 250 ms after stimulus onset.

Exp. 2. Exp. 2 was designed to rule out a potential sensory confound in Exp. 1, demonstrating that the ring of suppression observed in Exp. 1 was truly a result of the focusing of attention onto the target. In addition to being the focus of attention, the target was also the one red item in an array of homogeneous blue items. Pop-out items such as this receive enhanced sensory processing (36, 37), and this fact alone could cause the probe-elicited response to vary as a function of the distance between the probe and the pop-out item. Although it is unlikely that the center-surround pattern observed in Exp. 1 was entirely caused by sensory-sensory interactions of this sort, it is possible that the pattern was somewhat distorted by these interactions. Thus, Exp. 2 was conducted to measure the sensory–sensory interactions and subtract them away from the attention effect.

Exp. 2 contained two conditions that were variants of the procedure used in Exp. 1, tested in separate trial blocks. In both blocks, a rapid serial visual presentation (RSVP) stream of small characters at fixation was presented concurrently with the search frame and probe stimulus. In the attend-RSVP block, observers had to perform a demanding target detection task with the RSVP stimuli, and the search frames were irrelevant. This condition was designed to withdraw attention from the search frames, making it possible to evaluate the pure sensory effects of varying the target-probe distance. The observers' performance on this demanding task was quite good (94.2% correct) and did not vary between probe and no-probe trials (P = 0.6), confirming that observers were consistently focusing attention on the RSVP stream. In the attend-search block, observers ignored the RSVP stream and performed the search task as in Exp. 1. We predicted that the effects of target-probe distance would be small or nonexistent in the attend-RSVP condition, whereas the attend-search condition would replicate the pattern observed in Exp. 1.

Fig. 4a shows the mean size of the probe effect (FP-minus-FO trials) from 130–150 ms for Exp. 2. Black bars show the effect in the attend-search condition, and gray bars show the effect in the attend-RSVP task. When search frames were ignored, the probe response exhibited a slight reduction when the red target C appeared at or near the probe's location. However, pairwise comparisons between PD1 and the adjacent target locations (PD0 and PD2) revealed no significant effect (both Fs <1). In contrast, when observers performed the search task, the probe response profile replicated the pattern observed in Exp. 1, with a suppression at PD1 compared with the other target-probe distances. This finding was statistically confirmed by significant differences between PD0 and PD1 (F_1,7 = 18.6, P < 0.005), between PD2 and PD1 (F_1,7 = 12.6, P < 0.01), and between PD4 and PD1 (F_1,7 = 7.1, P < 0.05).

Fig. 4b shows the difference in response between the attend-search and attend-RSVP conditions, which subtracts the pure sensory response from the attention effect. These difference values were evaluated statistically. A significant enhancement was observed at the target location (PD0, F_1,7 = 18.8, P < 0.005), and a significant suppression was observed when the target was adjacent to the probed location (PD1, F_1,7 = 9.2, P < 0.05). Although the difference scores at PD2–PD4 were all positive, they were not statistically significant.

These results demonstrate that attention causes a true suppression of sensory-evoked activity in a narrow ring surrounding the attended location. That is, when the search target was attended, the response at PD1 was suppressed even below the response observed at PD1 when attention was directed to a demanding foveal task, which provides a very conservative baseline.

Discussion

In this study we investigated the spatial profile of cortical activity surrounding the focus of attention in an array containing a target surrounded by distractors. Cortical responsiveness was probed with a task-irrelevant luminance increment at a fixed spatial position while the distance of the focus of attention was varied relative to the probe. Instead of a simple monotonic gradient, the observed activity profile was found to be shaped like a “Mexican hat,” with an enhanced probe response at the focus of attention, a suppressed response in a narrow zone surrounding the attended location, and a rebound at more distant location. A current source density analysis indicated that this center-surround profile is present in early-to-intermediate areas of visual cortex. Because we probed only along a single dimension, we did not show suppression around a complete ring. However, there is no reason to believe that attention-related suppression would be present above and below a target but not to the left and right of the target, and a ring-shaped region of inhibition is therefore the most likely explanation for the observed results.

Previous single-unit, event-related potential (event-related potential) and ERMF studies have shown that occipito-temporal attention mechanisms are focused onto a pop-out target within 200 ms of the onset of the search array (17, 38–41), and previous event-related potential studies have shown that probes presented at 250 ms elicit enhanced occipito-temporal responses when presented at the target location compared with a location in the opposite visual field (32, 33). Prior behavioral and physiological studies of the spatial distribution of attention, however, have provided mixed results, either in favor of a simple spatial gradient (18–22), or consistent with an inhibitory zone (25–31, 42–44). The present study demonstrates unequivocally that attending to an object leads to a ring of inhibition. Previous failures to observe this pattern may reflect an inadequate spatial sampling or the use of experimental paradigms in which the target and distractors were not present simultaneously.

In addition, electrophysiological evidence for an inhibitory zone surrounding the focus of attention has been limited with respect to the timing of this effect. Slotnick et al. (30) report data from a probe paradigm in which subjects were asked to constantly focus onto a target location for ≈50 s while the probe's location varied relative to the target. Based on inverse dipole modeling of the event-related potential data, an inhibitory region surrounding the attended location was inferred under these conditions. However, these data do not indicate whether this inhibitory zone arises rapidly enough to represent a correlate of fast and flexible focusing, or instead, evolves gradually over several seconds. Related functional MRI studies addressing this issue were limited in an analogous way (14, 31), aside from the low temporal resolution inherent in this method. The present data now demonstrate that the Mexican hat profile builds up rapidly under conditions of fast and flexible attentional focusing, as required in visual search. The occurrence of the effect within 250 ms after search frame onset is therefore consistent with the time course of attention-based suppression effects observed at the cellular level in monkeys performing visual search (9), as well as with behavioral estimates of the attentional dwell time in visual search in humans (45, 46).

In view of recent evidence demonstrating retinotopically specific suppression effects during the anticipation of distractors (13), it is important to consider whether the use of a predictable probe location may have led subjects to develop a sustained inhibition at the probe's location (PD0), irrespective of the location of the target on a given trial. Indeed, RTs were slowed when the target appeared at the location where the probes were presented, even on probe-absent trials. However, this sustained inhibition would have had the same impact on the probe response irrespective of the location of the target, and it cannot therefore explain the observed variations in probe amplitude as a function of target-probe distance. It is also possible, although unlikely, that subjects specifically suppressed the PD0 location when they noticed that the target appeared at the PD1 location, anticipating that a probe might occur at the adjacent location. This seems like an implausible strategy, because the simultaneously presented distractors at the adjacent locations presumably interfered with perception of the target much more than the probe, which was presented 250 ms later. Indeed, the presence of a probe caused only a 4-ms slowing of RTs, and this slowing was independent of target-probe distance. Moreover, a ring of inhibition has been observed behaviorally even when distractor locations were unpredictable (28). Thus, the observed effects almost certainly reflect a temporary suppression that was triggered around the location of the target shortly after the appearance of the target.

The present results are consistent with the biased competition theory of attention of Desimone and Duncan (47). According to this theory, attention operates when multiple stimuli compete for access to neural representation, and this competition occurs when multiple stimuli fall within a neuron's receptive field. As a result, distractors within a receptive field are suppressed while attended stimuli are enhanced. How this mechanism operates within a hierarchical network of neurons is unspecified, but the general idea is consistent with our results.

A more detailed theory, the ST of Tsotsos (1, 23, 24) provides a network mechanism to accomplish such a biased competition. ST represents a computational account of attentional selection across the visual cortical processing hierarchy, based on realistic anatomical and physiological assumptions about the connectivity in the visual system. Although a number of computational models have proposed principles of cross-hierarchical processing in the visual system with functional explanations of how spatial focusing may be brought about (48–50), the ST directly predicts the center-surround pattern observed in this study as an emergent property of its architecture.

ST provides a solution for a paramount coding problem of the visual system that mainly arises from the massive upstream receptive field (RF) convergence in the visual processing hierarchy, which produces very large receptive fields at the top. That is, parts of a scene that fall into separate RFs at lower levels become inseparable within larger RFs at higher representational levels. Thus, focusing onto a small portion of a cluttered visual scene would be increasingly compromised due to the spatial confusion between separate scene parts. To achieve spatial selectivity, it would be necessary for the visual cortical hierarchy to keep the range of activation narrow across many if not all of its levels.

As illustrated in Fig. 4c, ST proposes that spatial selectivity is accomplished by a sequence of top-down propagating WTA mechanisms that attenuate irrelevant spatial representations (connections) iteratively from one hierarchical level to the next, thereby narrowing the range of cortical activation to match the extent of relevant details. Specifically, an initial feed-forward activation is triggered by the input and causes an activity distribution like an inverse pyramid crossing hierarchical levels. At the highest levels, global WTA processes select the strongest responding units, under the assumption that a strong response corresponds to the best match between stimulus and neural selectivity. Winning units from this WTA process start subsequent top-down WTA processes (pyramid-shaped), layer by layer, that prune away upstream connections that were activated by the initial feed-forward process, but lose the WTA competition and thus are not strong candidates for representing the target's location (Fig. 4c). Connections that survive the pruning process form the “pass zone,” or the pathways within which the selected stimulus is optimally processed (red in Fig. 4c). As an emergent property of this architecture, a surround of relative inhibition is produced around the pass zone representing the attended location.

This model accommodates two key observations: (i) it explains the relative enhancement of neural activity at attended locations, and how selectivity can be flexibly rerouted to favor new locations; and (ii) it explains the seeming paradox that attentional modulations in higher level areas appear earlier than in lower level areas (10, 51–53).

In sum, the present findings add to mounting evidence emphasizing that attentional focusing involves the simultaneous operation of neural enhancement and neural suppression (11, 12, 14, 15, 30). These findings also indicate that enhancement and suppression may cooperate in a spatially structured manner that optimizes noise reduction during visual object recognition, as predicted by ST. That is, the center-surround profile represents an activity distribution that is optimal to locally demarcate the target from non-target information, specifically attenuating inputs from nearby distractor items that would be at the largest risk to confuse target discrimination processes.

Materials and Methods

Stimuli and Task in Exp. 1. The stimuli and task are illustrated in Fig. 1. While fixating the center of the screen, observers searched for a red target C among eight blue distractor Cs presented at an isoeccentric distance from fixation (8° of visual angle) in the right lower quadrant (Fig. 1a). Each search frame was presented for 700 ms, followed by an interstimulus interval of 650–850 ms. Spacing between Cs was constant (1.35°), and each C subtended 0.8° of visual angle. The gap of each C varied randomly in position, and observers indicated the position of the gap in the target C (left or right) by pressing one of two buttons with the right hand. The target C appeared randomly at one of the nine possible stimulus locations (illustrated in Fig. 1c). On 50% of the trials, a white ring (the probe stimulus) was flashed around the center C for 50 ms; the probe onset occurred 250 ms after the search frame onset on FP trials. In the other 50% of trials, no probe was presented (FO trials, Fig. 1b). Because the probe position was constant and the target position varied, there were five target-to-probe distances, ranging from PD0 (target at probed location) through PD4 (target four items away from probe; see Fig. 1c). Each experimental session was separated into 10 runs lasting 6 min. During each run, 90 FP and 90 FO trials were presented, with 10 trials per probe-distance condition, amounting to a total of 100 FP and FO trials for each probe-distance condition throughout the complete session.

Stimuli and Task in Exp. 2. The stimuli and procedure of Exp. 2 were identical to those of Exp. 1, except for an additional RSVP stream of characters (0.8°) presented at fixation. Characters were randomly taken from a list of 10 uppercase items (A, E, I, K, L, N, T, V, Y, and O) and presented in a rapid temporal sequence with a new character appearing every 80 ms. The RSVP stream started 100 ms before the onset of the search frame and continued for 800 ms, ending 100 ms after search frame offset with the presentation of a question mark. In 30% of the trials, a target character (X) appeared in the RSVP stream; in the other trials, the target was absent. In RSVP blocks, the peripheral stimuli were irrelevant, and observers were instructed to attend the RSVP stream to detect the presence of the target character. In search blocks, observers were instructed to perform the search task as in Exp. 1 and to ignore the RSVP stream at fixation. Each observer performed five RSVP blocks and five search blocks.

Data Acquisition and Analysis. The magnetoencephalographic (MEG) signal was recorded by using a BTi Magnes 2500 whole-head MEG system (Biomagnetic Technology, San Diego) with 148 magnetometers. Data processing including offline noise reduction, artifact rejection, and low-pass filtering (DC-50Hz), following a protocol described in detail in refs. 38 and 39.

Average ERMF waveforms were computed for each subject, time-locked to probe onset, with a 250-ms prestimulus baseline interval. Separate averages were computed for each probe-distance condition (PD0–PD4) of FO and FP trials. To isolate the ERMF response elicited by the probe (called the probe–related response) from the overlapping response elicited by the search array, FO waveforms were subtracted from FP waveforms (FP-minus-FO difference) of trials with the target C at identical positions. The size of the probe-related response was quantified in each observer as the mean amplitude of the ERMF difference between the efflux- and influx-maximum between 130 and 150 ms, relative to the baseline. Sensor sites showing the efflux- and influx-maximum varied between subjects but were identical for all probe-target distances for a given subject. The time-range of analysis was determined by a one-way ANOVA with a factor of probe-target distance (five levels) that was computed for subsequent time samples from 100–200 ms. This analysis revealed a significant effect (P < 0.05) between 125 and 150 ms.

For source localization, current SDEs were computed by using a minimum norm least squares method as implemented in the curry 4.0 neuroimaging software (Neurosoft, El Paso, TX). Computations were performed on the basis of realistic anatomical constraints derived from 3D-surface segmentations of the cortical surface (source compartment) and the cerebrospinal fluid space (volume conductor) of the standard Montreal Neurological Institute (MNI) brain (see ref. 39 for details). To maximize signal-to-noise ratio, the localization was performed by using data averaged across all observers. This averaging may blur the SDE distribution somewhat, but the goal of the localization was to estimate the general location of the activity, not to make fine discriminations between nearby brain regions.

Subjects. Twelve observers (mean age, 23.4 y) took part in Exp. 1; 8 observers (mean age, 21.9 y) took part in Exp. 2. All observers were neurologically normal students of the Otto-von-Guericke-University, gave informed consent, and were paid for participation. Both experiments were approved by the ethics committee of the Otto-von-Guericke-University.