Partial Occlusion Modulates Contour-Based Shape Encoding in Primate Area V4

Transcription

1 12 The Journal of Neurosiene, Marh 16, 11 31(11):12 24 Behavioral/Systems/Cognitive Partial Olusion Moulates Contour-Base Shape Enoing in Primate Area V4 Brittany N. Bushnell, Philip J. Haring, Yoshito Kosai, an Anitha Pasupathy Department of Biologial Struture an National Primate Researh Center, University of Washington, Seattle, Washington Past stuies of shape oing in visual ortial area V4 have emonstrate that neurons an aurately represent isolate shapes in terms of their omponent ontour features. However, rih natural senes ontain many partially olue objets, whih have aiental ontours at the juntion between the olue an oluing objets. These ontours o not represent the true shape of the olue objet an are known to be pereptually isounte. To isover whether V4 neurons ifferentially enoe aiental ontours, we stuie the responses of single neurons in fixating monkeys to omplex shapes an ontextual stimuli presente either in isolation or ajoining eah other to provie a perept of partial olusion. Responses to preferre ontours were suppresse when the ajoining ontext renere those ontours aiental. The observe suppression was reverse when the partial olusion perept was ompromise by introuing a small gap between the omponent stimuli. Control experiments emonstrate that these results likely epen on ontour geometry at T-juntions an annot be attribute to mehanisms base solely on loal olor/luminane ontrast, spatial proximity of stimuli, or the spatial frequeny ontent of images. Our finings provie novel insights into how olue objets, whih are funamental to omplex visual senes, are enoe in area V4. They also raise the possibility that the weakene enoing of aiental ontours at the juntion between objets oul mark the first step of image segmentation along the ventral visual pathway. Reeive Sept. 6, ; revise Jan., 11; aepte Jan. 13, 11. This work was supporte by National Eye Institute Grant R1EY18839, the Whitehall Founation, University of Washington Vision Core Grant PEY17, an National Center for Researh Resoures Grant RR166. We thank Wyeth Bair, Jalal Baruni, Greg Horwitz, an Yasmine El-Shamayleh for helpful isussions an omments on this manusript. Jalal Baruni, Mari Kalif, an National Primate Researh Center Bioengineering provie tehnial support. Corresponene shoul be aresse to Anitha Pasupathy, Department of Biologial Struture an National Primate Researh Center, University of Washington, 1959 Paifi Street N.E., HSB G-5, UW Mailbox 3574, Seattle, WA pasupat@u.washington.eu. DOI:.1523/JNEUROSCI Copyright 11 the authors /11/ $15./ Introution When the three-imensional worl asts a two-imensional image on the retina, objets loser to the viewer an partially olue objets that are farther away. Uner suh onitions, the retinal images of olue objets are istorte an reognition must rely not only on the visual features of the objet but also on information erive from neighboring objets an sene ontext. Beause few physiologial stuies (Kovás al., 1995; Missal et al., 1997; Sugita, 1999; Bakin et al., ; Zhou et al., ; Kourtzi an Kanwisher, 1; Lerner et al., 4; e Wit et al., 6; Murray et al., 6; Raushenberger et al., 6; Fallah et al., 7) have investigate how neighboring or overlapping stimuli moulate responses of neurons in the ventral shape proessing pathway, we know little about how, where, or when information from sene ontext is inorporate into neural signals (for review, see Albright an Stoner, 2) an how these signals ultimately unerlie image segmentation an sene pereption. Here we quantify how responses of neurons in visual ortial area V4, an intermeiate stage along the ventral pathway, are moulate by ontextual information relate to olusion an propose how the observe moulations might ontribute to image segmentation. Previously, we emonstrate that V4 responses faithfully enoe omplex shapes in terms of their omponent ontour features (Pasupathy an Connor, 1, 2), suh that reognition oul be suessfully base on the V4 population representation. However, when the visual sene ontains partially olue objets, aiental ontour features, also alle extrinsi ontours (Nakayama et al., 1989), are forme at the resulting T-like juntions (angles an in Fig. 1a). Aiental ontours o not provie information about the true shape of the omponent objets an are less salient than real ontour features (see Fig. 1a,b, ompare an ), an shape jugments base on aiental ontours are slower than those base on real ontours (Gerbino an Salmaso, 1987; Rensink an Enns, 1998). Thus, we investigate whether V4 responses enoe real an aiental ontours ifferentially. For example, a neuron tune to sharp onvexities will respon preferentially to the isolate resent in Figure 1b. When the resent is ajoine by a irle (see Fig. 1a), the sharp onvexities an the onavity are pereptually evalue an the re shape is pereive as an ellipse. Do V4 responses that enoe the sharp onvexities an the intervening onavity reflet this pereptual evaluation? The broa onvexity of the resent in Figure 1a remains a real an salient ontour even in the presene of the ajoining stimulus. Is this also reflete in the responses of neurons that enoe this non-aiental broa onvexity? We aress these questions by omparing the responses to shapes presente in isolation an in the presene of ontextual stimuli that suggest partial olusion. Our results iniate that V4 responses ifferentially enoe real an aiental ontour features, an the ifferenes appear soon after response onset.

2 Bushnell et al. Partial Olusion an Shape Enoing in V4 J. Neurosi., Marh 16, 11 31(11): Contextual Stimuli Primary Shapes Combination Stimuli Real Aiental a θ φ Aiental θ φ Real Figure1. Illustrationofaientalontourfeaturesanstimulusset.a,Angles an areaientalontourfeaturesforme at the T-juntions between the oluing (blue) an olue (re) shapes. These angles are physially iential to those in b but are pereptually less salient., Stimulus set use in the primary experiment. Primary shapes (mile row) were presente in the preferreoloreitherinisolationorajoinebyontextualstimuli(toprow) inthenonpreferreolor. Thesharponvexitiesofthe primary shape are aiental ontours in the ombination stimuli (bottom row), but the broa onvexities remain real ontours. Eah ell was teste with a subset of orresponing primary, ontextual, an ombination stimuli. Asterisks mark the stanar subset that was use when preliminary shape-tuning tests were not onute. Control experiments suggest that our finings likely epen on ontour geometry at T-juntions an annot be attribute to mehanisms base solely on loal olor/luminane ontrast, spatial proximity of stimuli, or the spatial frequeny ontent of images. Materials an Methos Animals an surgery. Two rhesus monkeys (Maaa mulatta, 6 kg female an 7 kg male) were surgially implante with ustom-built hea posts attahe to the skull with orthopei srews. Animals were seate in front of a omputer monitor at a istane of 57 m an were traine to fixate a.1 white ot within.5.75 of visual angle. Eye position was monitore using a Hz infrare eye-traking system (Eyelink ; SR Researh). Stimulus presentation an animal behavior were ontrolle by ustom software (PYPE) originally evelope in the Gallant (University of California, Berkeley, Berkeley, CA) an Mazer (Yale University, New Haven, CT) laboratories. One animals were traine on the fixation task, a low-profile titanium ring (4 mm in height) that serve as the base of the reoring hamber was attahe to the skull with orthopei srews. The ring plaement, base on strutural magneti resonane imaging sans, inlue both the lunate an superior temporal suli. Skin was pulle over the ring an allowe to heal. A raniotomy was performe in a subsequent surgery, an a plasti reoring hamber (inner iameter of 19 mm) was attahe to the titanium ring with set srews. All animal proeures onforme to National Institutes of Health guielines an were approve by the Institutional Animal Care an Use Committee at the University of Washington. Data olletion. During eah reoring session, a single urapunturing miroeletroe (FHC In.), 25 m in iameter, was lowere into the brain using an eight-hannel aute mirorive system (Gray Matter Researh) whih allowe iniviual ajustment of eletroes via omputer ontrol of miniature stepper motors. Triggere waveforms from the eletroe were amplifie an filtere, an single-neuron ativity was isolate using a 16-hannel spike sorting system (Plexon Systems). During our initial reoring sessions, eletroe penetrations spanne the anteroposterior an meiolateral (ML) extent of the raniotomy to asertain the loation of the lunate sulus beause V4 oupies the prelunate gyrus an ajoining sulal banks. We ifferentiate between V4 an neighboring V2 neurons primarily base on reeptive fiel (RF) size (V4 RFs are larger that V2 RFs), RF loation (at similar ML extent, V4 RFs are more eentri than V2 RFs; see below for V4 RF size/loation information), an physiologial harateristis (on average, V2 responses to oriente bars are far more robust than those of V4 neurons). Visual stimuli. Visual stimuli were presente on a athoe ray tube monitor (.6.5 m; 97 Hz frame rate; 16 pixels) against a gray bakgroun of mean luminane of 5.4 /m 2. Stimulus onset an b offset were base on photoioe etetion of synhronize pulses in the lower left orner of the monitor. Eah isolate unit was initially haraterize with rifting or flashing bars, ellipses, gratings, an a variety of other shapes (primary an ontextual shapes illustrate in Fig. 1) uner the experimenter s ontrol. This haraterization ientifie an initial preferre stimulus (shape, olor, orientation) an an approximate RF loation. This was followe by an automate RF mapping proeure that presente the initial preferre stimulus in a ensely sample gri that spanne twie the han-mappe RF area. The refine RF enter was base on a two-imensional Gaussian fit to the ata. To ientify preferre an nonpreferre olors, we haraterize olor tuning with 25 olors presente at four ifferent luminanes (2.7, 5.4, 8.1, an 12.1 /m 2 ). The 25 olors provie an approximately uniform sampling of the CIE (for International Commission on Illumination) olor spae. For ells in our ataset, the etaile olor-tuning haraterization was bypasse in the interest of time, an the preferre an nonpreferre olors were hosen base on the initial haraterization uner experimenter s ontrol. For most neurons stuie, a thir preliminary test of shape tuning was onute using shapes (Fig. 1, subset shown in top row), presente at eight orientations, separate by 45 intervals, in the preferre olor of the ell. The preferre ontour feature was ientifie using previously esribe analytial proeures (Pasupathy an Connor, 1). For the remaining neurons, we proeee iretly to the primary experiment after olortuning haraterization beause results from the shape-tuning test were not neessary for onuting the primary experiment. Shape tuning for all neurons was haraterize base on responses to primary shapes an irles (see below). Primary experiment. To quantify how V4 responses are moulate by ontextual stimuli, we measure responses of neurons to a subset of the primary shapes shown in Figure 1 (mile row) presente either in isolation or ajoine by the orresponing ontextual stimulus. The ontextual stimuli (Fig. 1, top row) are shapes that were use previously to stuy V4 (Pasupathy an Connor, 1). The primary shapes were esigne suh that, when presente ajoining the orresponing ontextual stimulus, the perept is that of a partially olue irle. We have subjetively verifie this for the range of eentriities an stimulus sizes use in this stuy (see below). Furthermore, similar psyhophysial results on shape-mathing experiments base on aiental ontours at the enter of gaze an more eentri loations (V. Le an A. Pasupathy, unpublishe results) suggest pereptual similarity at these loations. Geometrially, primary shapes were onstrute by removing the intersetion between the ontextual stimulus an a irle. The full irles were also a part of the stimulus set. Of the 167 ells that unerwent the preliminary haraterization, we onute the primary experiment on 129 well-isolate V4 neurons that showe moerate olor an shape tuning uring preliminary tests. The remaining ells (38 of 169) that we i not pursue beyon the preliminary haraterization showe stronger responses to oriente bars or sinusoial gratings than to any other stimuli in our set an were not tune for speifi ontour features. This seletion allowe us to fous on ontour-tune neurons an investigate how their responses were moulate by an ajoining ontextual stimulus. For eah ell, we hose four to seven of the primary/ontext/ombination sets base on results from the preliminary shape test. Speifially, we hose ontextual stimuli that i not reliably rive the ell, to ensure that they only provie a moulatory influene. When the automate shapetuning test was not onute, we use a stanar set of five stimuli (Fig. 1, asterisks). Eah stimulus was presente at eight orientations separate by 45 intervals (Fig. 2). The primary shapes an irles were presente in the preferre olor an ontextual stimuli in the nonpreferre olor. We hose preferre an nonpreferre olors at the same luminane, both

3 14 J. Neurosi., Marh 16, 11 31(11):12 24 Bushnell et al. Partial Olusion an Shape Enoing in V4 a Primary shapes b Context stimuli Combination Swappe loation e Loal ontrast f Contrast normalization g Spatial separation h Tri-olor juntion i Cirles Figure 2. Sample stimuli for the primary an ontrol experiments. a, i, Example set of primary shapes (a), ontextual stimuli (b), ombination stimuli (), an irles (i), use in the primary experiment. Rows epit the five stanar shapes (asterisks in Fig. 1), an olumns epit the eight stimulus orientations teste. In all figures, re represents the preferre olor an blue the nonpreferre olor of the ell uner stuy. h, Example set of stimuli for the various ontrol experiments., The swappe loation ontrol investigates whether the observe effets are simply attributabletothepreseneofaseonstimulusoranonpreferreolorintherfbyalteringthespatialrelationshipoftheprimaryanontextualstimuliwithintherf(ompareorresponingions in an ). e, To etermine whether the observe results an be attribute to ifferenes in loal olor or luminane ontrast, on the loal olor ontrast ontrol, the primary shape stimuli are presente against a bakgroun of the nonpreferre olor. f, Contrast normalization ontrol. The preferre (re) an nonpreferre (blue) olors are hosen at ifferent luminanes on this ontrol to testwhethernormalizationmehanismsanexplaintheobserveresults. g, Spatialseparation. Thisexperimenttestswhetherseparationoftheprimaryanontextualstimulibyasmallgap, whih reverses the partial olusion perept, reverses the observe effets also. h, The triolor juntion ontrol investigates whether the spatial proximity of the preferre an nonpreferre olor image features is suffiient to proue the observe effets or if the T-juntion, implying partial olusion, is neessary. For aitional etails, see Results. either brighter or arker than the bakgroun. This prevente one of the two stimuli from automatially attrating attention an thereby prouing attention-epenent response moulations. To verify that the observe effets generalize to ases in whih there was a luminane ontrast between the primary shapes an ontextual stimuli, in 32 ells, we hose preferre an nonpreferre olors at ifferent luminanes (Fig. 2f) suh that aiental ontours were efine by at least a % ontrast. In 15 of the 32 ells, the luminane of the ontextual stimuli an primary shapes strale the bakgroun luminane, making the aiental ontours the highest ontrast ontours in the stimulus isplay. In our ataset, RF eentriities range from 1 an 12, with a meian of 5.4. Stimulus size was sale with eentriity suh that all parts of all stimuli were within the estimate RF area (estimate RF iameter RF eentriity base on ata from Gattass et al., 1988). In our ata, the relationship between RF eentriity an the SD ( )ofthe best-fitting Gaussian (see above) an be apture by the following: RF eentriity. Thus, if the RF bounary is efine as the point at whih responses roppe to half the peak value, then the RF iameter RF eentriity. This is omparable with the equation base on Gattass et al. (1988) above. Combination stimuli were entere within the RF. Primary an ontextual stimuli oupie iential positions in the RF when presente in isolation an as part of the ombination stimuli. The primary stimulus set inlues a restrite range of aiental ontour features all aiental ontour features are sharp, aute onvexities ajoine by onavities; none are broa onvexities. This is beause the oluing ontours of all ontextual stimuli were onvex projetions, an none were onave inentations. We hose onvex oluing ontours for two reasons. First, they are known to provie a more vivi olusion perept than onave oluing ontours (MDermott an Aelson, 4), an seon, beause there is a strong bias for enoing shapes in terms of sharp onvexities in V4 (Pasupathy an Connor, 1999, 1), our stimulus hoie allowe us to stuy the majority of neurons in etail. In a future stuy, onave oluing ontours will be inlue along with binoular ues to enhane the olusion perept. Eah trial began with the presentation of a fixation spot at the enter of the sreen. One fixation was aquire, four to six stimuli were presente in suession, eah for a uration of ms, separate by interstimulus intervals of ms. Stimuli were presente in pseuoranom orer 7 times. Thirty blank stimulus perios were interleave to alulate the spontaneous firing rate. Figure 2, a an i, epits a sample set of stimuli inlue in the primary experiment base on the five stanar primary shapes (Fig. 1, asterisks). To evaluate whether the observe effets oul be explaine by alternate hypotheses relate to the presene of a seon stimulus, loal olor, or luminane ontrast, we onute a series of ontrol experiments (Fig. 2 h) on a subset of neurons. The rationale behin these ontrols an their esign is explaine in Results. To ensure that the reporte finings are not simply attributable to nonstationarities in the measure responses, primary an ombination stimuli (Fig. 2a,) were also inlue in the ontrol experiments.

4 Bushnell et al. Partial Olusion an Shape Enoing in V4 J. Neurosi., Marh 16, 11 31(11): Data analysis. We ompute the mean response to eah stimulus by averaging the firing rate between stimulus onset an stimulus offset aross stimulus repetitions. Blank stimulus perios were use to erive spontaneous firing rates. Results presente here are base on mean responses without subtration of spontaneous rates; analyses base on mean responses after subtration of spontaneous ativity proue similar results. To etermine how ontext moulates responses to preferre ontours, we ompute a frational suppression inex, S ol, whih measures the suppression of mean ombination responses (R ombo ) relative to mean primary shape responses (R primary ) average aross all preferre primary shapes, i.e., S ol 1 n n R primary R ombo R primary, where n enotes the number of preferre primary shapes. To ientify the preferre primary shapes, we first normalize all responses to lie between. an 1., R norm (R R minimum )/(R maximum R minimum ) an then ientifie the primary shapes that evoke greater than half the maximum, i.e.,.5, in this normalize sale. It is essential to restrit our inex to stimuli eliiting greater than half-maximum responses to apture the suppression of preferre responses (see Fig. 3). Inluing nonpreferre stimuli woul have ilute the results for preferre stimuli, whih woul be unfavorable. Results were similar when the enominator of S ol was the sum of primary an ontextual responses (R ombo R ontext ) beause ontextual stimuli (in a nonpreferre olor) typially evoke weak responses. We preferre S ol to other measures suh as the slope of R ombo versus R primary beause the slopes an be zero even when preferre responses are not suppresse, for instane, when R ombo R primary for nonpreferre primary stimuli. To assess whether the observe suppression was simply attributable to the presene of a seon stimulus in the RF, we onute a swappe loation ontrol experiment in whih the loations of the primary an ontextual stimuli were interhange (Fig. 2). Thus, the exat same stimuli were in the RF but in a new spatial arrangement inonsistent with partial olusion. To evaluate whether suppression uner partial olusion was signifiantly ifferent from suppression for swappe ontrol stimuli (alulate using the same formula above), we use ranomization tests (Manly, 1997). For eah ell, we ompute a T statisti from suppression values alulate from iniviual trial responses [s ol (r primary r ombo )/r primary ; s swap (r primary r swap )/r primary ]. Then, response rates for iniviual stimulus repetitions were ranomly permute between the ombination an swappe loation ategories while maintaining stimulus ientity, an the test statisti was realulate. This proeure was repeate, times to onstrut the null istribution. Ranomization tests were also use to assess statistial signifiane of orrelation values. To test whether previously propose moels base on average or maximum firing rates an preit the responses to ombination stimuli, for eah neuron we quantifie the average, [ (R primary R ontext )/2.] an maximum [ max(r primary, R ontext )] preitors for eah ombination stimulus teste. We ompute the orrelation oeffiient between the average an maximum preitors an the ombination responses to assess the linear relationship between these variables. For all ells, we assesse shape tuning base on responses to primary shapes an irles using previously esribe analytial methos (Pasupathy an Connor, 1). Briefly, eah stimulus was represente as eight orere pairs of urvature angular position. Curvature values range from 1 (eep onavity) to 1. (sharp onvexity). Nonlinear least-squares methos were use to ientify the two-imensional Gaussian funtion in the urvature angular position spae, efine by two means ( urv, ), two SDs ( urv, ), an an amplitue, that best preite the observe responses; orrelation oeffiient between observe an preite responses quantifie gooness of fit (gof). We use two methos to measure lateny of suppression. First, to failitate iret omparison, we use the half-maximum lateny metho use by Zhou et al. () in V2. We onstrute population histograms for the preferre primary responses an ombination responses base on normalize responses of 76 neurons that showe statistially signifiant frational suppression. We then ompute the ifferene histogram by subtrating the ombination histogram from the primary histogram (see Fig. 11,e). The time from stimulus onset to half-maximum of the ifferene histogram measures the lateny of suppression. For eah neuron, we also quantifie the onset lateny for the responses of preferre primary an ombination stimuli by measuring the time from stimulus onset at whih the response exeee the mean baseline response by 3 SDs. Mean an SD of the baseline was base on the 75 ms perio before stimulus onset. We then quantifie suppression onset lateny for single neurons by onstruting a ifferene histogram in 5 ms bins between the preferre primary an ombination responses an then measuring the time, relative to stimulus onset, at whih the ifferene histogram exeee the mean baseline ifferene by 3 SDs. The lateny of suppression relative to response onset is therefore given by the ifferene between the lateny of suppression from stimulus onset an the visual response lateny of preferre primary responses. Results Enoing of aiental ontours We stuie the responses of 129 V4 neurons from two monkeys to a set of primary shapes an orresponing ontextual stimuli presente either in isolation or ajoining eah other. When the two shapes are ajoine, the ombination stimuli are likely to be pereive as irles partially olue by the ontextual shapes. This is attributable to two strong preitors of olusion: the urvature isontinuities at the T-juntions an the interruption in the outline of a familiar form (a irle) (Chapanis an MCleary, 1953). In this onition, the sharp onvexities of the primary shape are pereive as aiental ontour features (Fig. 1) (see Materials an Methos). Many V4 neurons that enoe sharp onvexities showe a ifferene in response when the preferre primary shape was ajoine by a ontextual stimulus versus when in isolation. An example is shown in Figure 3. In the preliminary shape haraterization, this neuron respone preferentially to stimuli with a sharp onvexity at the bottom of the shape (best-fit Gaussian urv.92; 292, gof.86). Primary stimuli at 225, 27, an 315, whih have this preferre feature, evoke strong responses from the ell (Fig. 3a). Cirles (Fig. 3f), primary shapes at other orientations that have broa onvexities at the bottom of the shape, an ontextual stimuli (Fig. 3b) (in the nonpreferre olor) evoke weak responses. When the primary shapes were ajoine by the orresponing ontextual stimuli, responses to preferre primary shapes were uniformly suppresse (Fig. 3). Figure 3 illustrates the marke suppression of preferre primary shape responses: all points in the right half of the plot (primary stimuli that evoke strong responses) lie in the bottom right orner, resulting in a poor orrelation between primary an ombination responses (r primary_vs_ombo.6). The ifferene between the orientation-tuning funtions for the resent primary shape, in isolation (Fig. 3e, re) versus in ombination (Fig. 3e, magenta), also illustrates the suppression of preferre responses in the partial olusion ontext. Differenes in average response rates are not simply attributable to inrease variability in the responses to ombination stimuli (Fig. 3g i). Previous results in V4 an inferotemporal (IT) ortex suggest that, epening on the stimulus an experimental methos, responses of neurons to multiple nonoverlapping stimuli an range from the maximum to the average of the omponent stimuli (Reynols et al., 1999; Gawne an Martin, 2; Zoolan et al., 5). When the two omponent stimuli are spatially ajoine, neither the maximum nor the average is well orrelate with the ombination responses (r max_vs_ombo.3; r average_vs_ombo.2).

5 16 J. Neurosi., Marh 16, 11 31(11):12 24 Bushnell et al. Partial Olusion an Shape Enoing in V4 a Primary shapes b Context stimuli Combination Combination response(spikes/se) Spikes/se f Cirles e 36 Primary response (spikes/se) Stimulus orientation spikes/se spikes/se g h i Primary shapes at 27 Combination at 27 Primary shapes Combination stimuli Time from stimulus onset (ms) Figure 3. Suppression of preferre responses uner partial olusion ontext. a, Average responses of an example neuron to four primary shapes (rows), ontext stimuli, an ombination stimuli presente at eight orientations (olumns) are shown in graysale. Blue bars in the bottom right orner of eah ion iniate SEM. a, Primary shapes with a sharp onvexity at the bottom of the shape( ) evoke strong responses from this ell. b, Contextual stimuli presente in the nonpreferre olor evoke weak responses., Preferre primary shape responses(a; ) were strongly suppresse in the presene of orresponing ontextual stimuli., Primary shape responses (as in a, x-axis) an ombination responses (as in, y-axis) were poorly orrelate: r.6. e, Stimulus orientation (x-axis) versus average responses (y-axis) for the primary shape, ontextual stimulus, an their ombination in the top row of a, b, an, respetively. Error bars iniatesem. Thestrongorientationtuningfortheresentprimaryshapeisabsentinthepreseneoftheontextualstimulus. f, Responsestoirleswereuniformlyweak. g i, Single-trialresponses an peristimulus time histograms for primary an ombination stimuli. Rasters an peristimulus time histograms, base on repetitions eah, of four primary stimuli presente in isolation (g) or in ombination with the orresponing ontextual stimuli (h). The four stimuli shown here are the primary shapes an ombination shapes at 27 orientation in a an, respetively. Gray lines show SEM. To test whether the observe suppression is simply attributable to the presene of a seon stimulus or a nonpreferre olor in the RF, we onute a swappe loation ontrol experiment in whih the relative positions of the primary an ontext stimuli were interhange (ompare orresponing stimuli in Fig. 2,). In this onfiguration, the same two stimuli were presente but there was no suggestion of partial olusion. Responses to the primary shapes an ontextual stimuli in the new positions (Fig. 4a,b) were similar to those in Figure 3, a an b. When the two stimuli were simultaneously presente in the new positions, however, responses were ramatially ifferent from those in Figure 3: no suppression of preferre responses was evient when the stimuli were not spatially ajoine (Fig. 4). This lak of suppression is apture by the strong orrelation between responses to primary shapes an the swappe loation a Primary shapes Swappe loation response (spikes/se) stimuli (Fig. 4)(r.81) an the preservation of orientation tuning for primary shapes in the swappe loation ontrol (Fig. 4e). Finally, in keeping with previous finings (Reynols et al., 1999; Gawne an Martin, 2; Zoolan et al., 5), responses to the simultaneous presentation of stimuli in this spatial onfiguration were well orrelate with both the maximum an the average of responses to the omponent stimuli (r max_vs_swappe loation.76; r average_vs_swappe loation.69; it is iffiult b Context stimuli Spikes/se 36 Primary response (spikes/se) Stimulus orientation e Swappe loation Figure 4. Results from the swappe loation ontrol experiment for the neuron in Figure 3. To test whether suppression observe in Figure 3 was simply attributable to the presene of a seon stimulus or nonpreferre olor in the RF, primary shapes an ontext stimuli werepresenteatswappeloationsintherf. AllotheronventionsareasinFigure3. ResponsesinaanbaresimilartothoseinFigure 3, a an b., Unlike in Figure 3, no suppression of preferre responses is observe for swappe loation ombination stimuli., Primary shape(x-axis)anombinationresponses(y-axis)intheswappeposition(asinaan,respetively)arestronglyorrelate(r.81). e, Orientation-tuning funtions (onventions as in Fig. 3e) for stimuli in the top row of a. to ifferentiate between these moels beause of the uniformly low responses to ontext stimuli in this stuy). In summary, when the preferre ontour was ajoine by a ontextual stimulus, the responses of this neuron were strongly suppresse. Results from the swappe loation ontrol iniate that the suppression annot be attribute simply to the presene of a seon stimulus, a nonpreferre olor, or other normalization mehanisms that are insensitive to the preise spatial arrangement of stimuli. spikes/se

6 Bushnell et al. Partial Olusion an Shape Enoing in V4 J. Neurosi., Marh 16, 11 31(11): Combination response (spikes/se) a Primary shapes b Context stimuli Spikes/se 36 Primary response (spikes/se) Responses to broa onvex ontours In our stimulus esign, the broa onvexities (Fig. 1a, resent) were pereive as a real ontour in both isolation an ombination, i.e., the aition of the ontextual stimulus oes not hange the interpretation of this ontour. This is beause none of our ontextual stimuli ha a onave oluing ontour that physially ajoine the broa onvexity (see Materials an Methos). Responses of the example neuron in Figure 5 mirror this onstany in pereption of the broa onvexity. This neuron respone preferentially to shapes with a broa onvexity at the bottom right orner of the shape ( urv.47; 3, gof.56) as emonstrate by the strong responses to irles (Fig. 5f), primary shapes at orientations from (Fig. 5a), an ontextual stimuli with this preferre ontour (Fig. 5b, top an bottom rows). Responses to the ombination stimuli (Fig. 5) were similar on average to the responses to primary shapes in isolation (Fig. 5)(r primary_vs_ombo.62). In fat, responses to ombination stimuli were marginally higher than for orresponing primary shapes (ranomization paire t test, p.1). Thus, unlike the example in Figure 3, the responses to this neuron to the preferre primary shapes were not suppresse even in the presene of the ontextual stimulus. Population results Figure 6 shows the results aross the population of 129 neurons. For eah neuron, the normalize irle response is plotte along the x-axis in Figure 6a. Cells that respone preferentially to broa onvexities lie near 1. as a result of strong responses to irles, whereas ells seletive for sharp onvexities or onavities respone weakly to irles an therefore lie to the left of 1.. The frational suppression inex, S ol, whih measures the suppression assoiate with the partial olusion onfiguration (see Materials an Methos), is plotte along the y-axis. Aross our e Combination f Cirles Figure 5. Responses of an example neuron enoing broa onvexities. All onventions as in Figure 3. a, Average responses to five primary shapes (top an bottom rows are iential). This neuron respone preferentially to primary shapes with a broa onvexity in the bottom right of the shape ( ). Shapes at other orientations evoke weak responses. b, Responses to ontextual stimuli. This ell i not exhibit strong tuning for olor, an some ontextual stimuli in the nonpreferre olor evoke moerate responses (top an bottom rows), refleting preferene for broa onvexities at the bottom right of the shape., Responses to ombination stimuli are similar to primary shapes in isolation(shown in a)., Primary(shown in a) an ombination (shown in ) responses are plotte along the x- an y-axes respetively. Responses are strongly orrelate (r.62). e, Orientation-tuning funtions for the resent primary shape (top row of a) an the orresponing ombination stimuli (top row of ) are very similar. f, Strong responses to irles verify preferene for broa onvexity. spikes/se population, there was a statistially signifiant negative orrelation (r.56, p.1) between normalize responses to irles an frational suppression: as the response to irles inrease, frational suppression erease. Neurons that were most seletive for sharp onvexities an onavities (that lie at the left extreme of the plot) showe strongest suppression. Neurons that were seletive for broa onvexities showe a broa range of suppression values, inluing a few (nine ells) that showe negative frational suppression values implying greater responses to ombination stimuli ompare with the primary stimuli (as in Fig. 5). On average, however, neurons seletive for broa onvexities showe less suppression than those seletive for sharp onvexities or onavities. Thus, aross our population of V4 neurons, responses to sharp onvexities renere aiental by ajoining ontextual stimuli were strongly suppresse, whereas responses to broa onvexities, whih remaine real ontours, were not. To further relate the observe suppression to the preferre ontour feature an to get a measure of the relative strength with whih the various ontours are enoe, Figure 6b shows average frational suppression as a funtion of the urvature-tuning peaks of neurons. Tuning peak, urv,ofthe best-fitting Gaussian is plotte along the x-axis, an the average frational suppression of neurons with tuning peaks that lie within the orresponing urvature bin is plotte along the y-axis. For eah neuron, the tuning peak was base on only shapes, an so the parameter estimates may not be aurate. However, the primary tren is learly evient: neurons tune to sharp onvexities an onavities showe stronger suppression than neurons tune to broa onvexities. This figure also suggests that, for ombination stimuli, neurons that respon best to shallow onvex urves, i.e., ells with tuning peaks ( urv )between. an.2, respone times as strong as neurons tune to sharp onavities ( urv 1.) or sharp onvexities ( urv 1.). One possibility is that neurons that are not olor tune are assoiate with high S ol values beause the intervening ontour between the primary an ontext stimuli (in the ombination onition) is efine only by a olor ontrast an thus essentially invisible to a ell that is not olor seletive. To explore this, Figure 6 shows the relationship between olor tuning an the observe suppression effets aross 98 ells that unerwent both the primary experiment an etaile olor haraterization. Breath of olor tuning, given by the number of olors (of 25) that evoke greater than half-maximum responses, is plotte along the x-axis an S ol along the y-axis. Both weakly (right extreme of x-axis) an strongly (left extreme of x-axis) olor-tune neurons were assoiate with high values of suppression; speifially, a positive orrelation between breath of olor tuning an strength of suppression is not evient. Thus, the observe suppression annot be attribute to weak olor seletivity. For every neuron, we also measure the strength of olor seletivity as the ratio of the ifferene (R pref R nonpref ) over the sum (R pref R nonpref )ofre-

7 18 J. Neurosi., Marh 16, 11 31(11):12 24 Bushnell et al. Partial Olusion an Shape Enoing in V4 Frational suppression S ol Inreasing seletivity for sharp onvex an onave Normalize irle response a Breath of olor tuning Average S ol S swap Curvature tuning peak sponses, R pref an R nonpref, to a preferre stimulus presente in the preferre an nonpreferre olors use in the primary experiment, respetively. Aross ells, we foun that there was a small but signifiant positive orrelation (r.29, p.1) between the strength of olor seletivity an strength of suppression. This further onfirms that the observe suppression was not simply the result of weak olor seletivity. The swappe loation ontrol (isusse above; see Fig. 4) was onute on a subset of 45 neurons, an the results are illustrate in Figure 6. For a majority of ells (35 of 45), S ol (x-axis) was greater than S swap (y-axis), an this was espeially true for ells with low irle responses (ot size is inversely proportional b S ol Figure 6. Population results. a, Eah ot enotes ata from a ell. x-axis, Cirle response normalize by the maximum primary shape response. Cells that preferentially enoe broa onvexities lie lose to 1. on this axis, whereas ells that enoe sharp onvexities an onavities lie lose to.. y-axis, Frational suppression, whih quantifies the ifferene between primary shape an ombination stimulus responses for stimuli that eliite greater than halfmaximumresponse(seematerialsanmethos). Arossthe129neurons, thenumberofstimuli with responses greater than half-maximum range from (for 11 ells) to (mean of 11.4, meian of ). Thus, 118 ells are inlue in this figure. Magenta irles plot frational suppressionofexampleellsinfiguresenotebythenumbers; thesevalueswerebaseon12(fig. 3) an (Fig. 5) stimuli that exeee half-maximum. There was a negative orrelation (r.56)betweennormalizeirleresponseanfrationalsuppression:ellsseletiveforsharp onvexitiesoronavitiesexhibitestrongersuppressionthanthoseseletiveforbroaonvexities. b, Average frational suppression as a funtion of tuning peak along the urvature axis. Contour urvature is plotte along the x-axis: 1 represents eep onavities an 1 sharp onvexities. y-axis shows average frational suppression aross neurons with urvature-tuning peaks in the orresponing bin., Breath of olor tuning(x-axis) versus frational suppression, S ol ( y-axis). Breath of olor tuning is given by the number of olors that evoke greater than half the maximum responses at the most responsive luminane on the olor haraterization test (see Materials an Methos). Thus, ells with weaker olor tuning are assoiate with greater breath of tuning. The example ells in Figures 3 an 5 (labele) show equally broa olor tuning but very ifferent S ol values. Aross the population, there was a negative orrelation (r.42) between breath of olor tuning an frational suppression., x-axis, S ol, frationalsuppressionforombinationstimuli;y-axis,s swap,frationalsuppressionforswappe loation ontrol stimuli, i.e., for primary an ontext stimuli presente in the interhange spatial loations. Dot size is inversely proportional to normalize irle response. Cells with smaller irle response (larger ots) lie below the iagonal (S ol S swap ). Green ots enote ells for whih S ol was signifiantly greater than S swap. Cells with large irle response (small ots) lie along the iagonal (S ol S swap ) an lose to the origin (S ol an S swap are small in magnitue). For the ell labele 1 on the x-axis, S swap.8. to irle response). Cells with low irle responses (larger ots) lie below the iagonal, i.e., S ol S swap. Cells with higher irle responses (smaller ots) lie lose to the iagonal but also showe lower levels of suppression overall, i.e., S ol an S swap were both small. If we onsier ells with normalize irle responses.8, 26 of 29 showe S ol S swap (meian: S ol.47, S swap.24), an this ifferene was statistially signifiant on 19 of 26 ells (green ots, p.5). For ells with irle responses.8, frational suppression tene to be quite low (as in Fig. 5) (meian: S ol.6, S swap.16), an S swap was just as often higher (7 of 16) as it was lower than S ol. The utoff of.8 is arbitrary, but it helps to illustrate that, for ells assoiate with weaker responses to irles, not all of the suppression an be attribute to the presene of a seon stimulus in the RF. In summary, in keeping with previous stuies, most neurons in our population showe some suppression (S swap for 35 of 45 neurons) as a result of the presene of a seon stimulus in the RF. However, ells tune for sharp onvexities or onavities (lower normalize irle responses) exhibite aitional suppression (S ol S swap ) that was speifi to the spatial relationship between the primary an ontext stimuli (i.e., they neee to be spatially ajoine). In ontrast, for ells that enoe broa onvexities, the hypothesis that the observe suppression is simply attributable to spatially nonspeifi mehanisms annot be rejete. This tren is also apture by the statistially signifiant negative orrelation between normalize irle responses an S ol S swap (r.45; p.1). We next onsier several alternate hypotheses, not relate to partial olusion, to explain the observe results an elineate the geometri onstraints require for prouing the observe suppression. Loal olor an luminane ontrast First, we onsier the hypothesis that loal olor an luminane ontrast unerlie the observe suppression. In our stimulus esign, the olor an luminane ontrasts at the sharp onvexities an onavities were not iential in the primary shape an ombination stimulus onfigurations. For primary shapes, all ontours were surroune by the bakgroun olor an efine by luminane ontrasts ranging from 66 to 77% (for luminane etails, see Materials an Methos), but for ombination stimuli, the sharp onvexity an onavity were ajoine by the nonpreferre olor. Beause preferre an nonpreferre olors were hosen to be equiluminant (to avoi ifferential attentional alloation; see Materials an Methos), these ontours were efine by olor ontrast but not luminane ontrast (Fig. 7f). To test whether the observe suppression was attributable to ifferenes in loal olor an/or luminane ontrasts, we stuie responses of 16 neurons to primary shapes presente against a bakgroun of the nonpreferre olor of the ell (Fig. 2e). The sharp onvex an onave ontours of the loal ontrast ontrol stimuli were efine by the same olor an luminane ontrast as the aiental ontours in the ombination stimuli. Results for an example neuron are shown in Figure 7a e. This neuron showe suppression of preferre primary shape responses (Fig. 7a) when ajoine by ontextual stimuli (Fig. 7b)(S ol.77) but not when the primary shapes were presente against a nonpreferre bakgroun olor (Fig. 7) (frational suppression.3). Responses to loal ontrast ontrol stimuli were strongly orrelate with primary shape responses (r.95) (Fig. 7e) but only weakly orrelate with ombination responses (r.36) (Fig. 7). For 15 of 16 ells, responses to the loal ontrast ontrol stimuli were better orrelate with primary shapes (meian of r.53) than with ombination stimuli (meian of r.4), an, for 14 of

8 Bushnell et al. Partial Olusion an Shape Enoing in V4 J. Neurosi., Marh 16, 11 31(11): a Primary shapes b Combination Loal ontrast ontrol Combination response (spikes/se) Loal ontrast ontrol response (spikes/se) Primary response (spikes/se) e f Luminane ontrast = % Luminane ontrast = % Luminane ontrast = % Combination stimulus Primary shape Loal ontrast ontrol Figure7. Loalontrastontrolresultsforanexampleneuron. a, Responsestofiveprimaryshapes(rows). Thisneuronresponepreferentiallytoshapesat18-27 orientations. b, Responses to orresponing ombination stimuli. Responses to preferre primary shapes are suppresse in the presene of the ontextual stimuli., Responses to loal ontrast ontrol stimuli. When primary shapes were presente against a nonpreferre olor bakgroun (same olor as the ontextual stimuli), no suppression was observe, i.e., response patterns in a an are similar., Primary shape responses (shown in a, x-axis) an ombination responses (shown in b, y-axis) were poorly orrelate (r.3). e, Loal ontrast ontrol responses (shown in, y-axis) were strongly orrelate (r.77)withprimaryshaperesponses(x-axis).f,stimulusonfigurations.beausepreferreannonpreferreolorswerehosentobeequiluminant,therewasnoluminaneontrastarossthe intervening bounary for ombination stimuli. The orresponing bounary of the primary shape was efine by a non-zero luminane. For the loal ontrast ontrol stimuli, olor an luminane ontrasts aross the orresponing bounary were iential to the ombination stimuli. spikes/se 16 neurons, the orrelation between loal ontrast ontrol stimuli an ombination stimuli was not signifiantly ifferent than zero (ranomization test, p.5). These finings suggest that the suppression of preferre responses observe in the partial olusion ontext annot be explaine by hanges in olor or luminane ontrast in the viinity of the relevant ontour feature an that equiluminane between the primary an ontextual stimulus alone is not suffiient to proue the observe suppression. We next onsier the possibility that the observe suppression of equiluminant ontours may be the result of population-base normalization mehanisms that ome into play when the visual sene inlues high ontrast ontours. Speifially, strong responses aross the population to the high ontrast ontours (between shapes an bakgroun) in the ombination stimuli oul suppress the responses to the zero ontrast aiental ontours. If this were the ase, then suppression woul be weak or absent when the intervening ontour between primary an ontext shapes was also efine by a high luminane ontrast. To test this preition, in 32 ells, we hose preferre an nonpreferre olors at ifferent luminanes suh that the aiental ontours were efine by at least a % ontrast in luminane (Fig. 2f); for 15 ells, the aiental ontours were the highest ontrast ontours in the stimulus isplay (see Materials an Methos). Results from a neuron teste in this way are shown in Figure 8a e. Primary shapes at 9 evoke strong responses (Fig. 8a), an these preferre responses were suppresse when ajoine by the orresponing ontextual stimuli (Fig. 8). Frational suppression (S ol.69) an the poor orrelation between primary an ombination responses (r.3) were omparable with results in Figure 3. Population results aross all 32 ells (Fig. 8f) followe a similar tren to those in Figure 6a: the range an pattern of suppression values aross neurons was very similar to that observe with equiluminant preferre an nonpreferre olors. Thus, responses to aiental ontours were suppresse even when they were the highest ontrast ontours in the visual isplay. This, therefore, argues against the possibility that the observe suppression is simply the result of preferential enoing of high ontrast ontours in the visual sene. Spatial separation When a small gap is introue between the primary shape an ontextual stimuli (Fig. 2g), partial olusion is not pereive. To relate neural ativity to this pereptual transition, we investigate how the extent of suppression epene on the spatial separation between the primary shape an the orresponing ontextual stimulus. In 17 ells that showe strong suppression in the primary experiment (S ol.4), we stuie responses at three ifferent spatial separations:.1,.2, an.5 RF raius (the largest separation was.75 an 1.25 in two ells). The example neuron in Figure 9 showe strong suppression (S ol.52) of preferre ontour responses in the primary experiment. Suppression was markely erease (frational suppression,.13) when a very small istane (.1 RF raius) separate the primary an ontextual stimuli (Fig. 9, ompare, ). Frational suppression was further reue, on average, for larger separations (.8 at both.2 RF raius an.5 RF raius) (Fig. 9e,f). All 17 ells showe a similar statistially signifiant epenene of suppression on spatial separation (ranomization one-way ANOVA, p.5) (Fig. 9g), an 14 of 17 ells showe a negative orrelation between spatial separation an frational suppression, i.e., suppression erease with inreasing separation. Rate of hange of suppression as a funtion of spatial separation varie aross ells. At.1 RF raius, the erease in suppression was 22% on average. This is less ramati than the example in

9 J. Neurosi., Marh 16, 11 31(11):12 24 Bushnell et al. Partial Olusion an Shape Enoing in V4 a Primary shapes b Contextual stimuli Combination spikes/se Combination response (spikes/se) e f.8 Spikes/se Primary response (spikes/se) Normalize irle response Figure 8. Example ell an population results for primary shapes an ontextual stimuli presente at ifferent luminanes. a, Responses to primary shapes presente in a preferre olor (luminane, 2.7 /m 2 ; bakgroun luminane, 5.4 /m 2 ). This neuron respone preferentially to shapes with a sharp onvexity at the top right orner. b, Responses to ontextual stimuli presenteinanonpreferreolor(luminane,8.1/m 2 ).Moeratelystrongresponsestoafewontextstimulirefletpreferenesforshapeswithasharponvexityatthetopright.,Combination responses showe a strong suppression of preferre primary shape responses., Primary shape responses (as in a, x-axis) an ombination responses (as in b, y-axis) were poorly orrelate (r.3). e, Orientation tuning for top row of primary shapes, ontextual, an ombination stimuli. f, Population results for 32 neurons (blak ots) teste with preferre an nonpreferre olors at ifferent luminanes. x-axis, Normalize irle response; y-axis, frational suppression. Blue ots are results from Figure 6a (preferre an nonpreferre olors at the same luminane) replotte for omparison. Magenta ientifies example ell in a e. S ol.4 Figure 9 but perhaps not unlike pereption beause the spatial separation neee to ause a pereptual hange epens on the eentriity of the stimulus loation (Le an Pasupathy, unpublishe observation). For a few ells, suppression at.5 RF raius was greater than at.2 RF raius. This may be beause parts of the ontextual shapes stimulate the inhibitory surroun for.5 RF raius. Aitional experiments are require to preisely haraterize the spatial profile of the suppression an to relate behavioral performane, as a funtion of spatial separation an stimulus eentriity, to neurophysiology. Triolor juntions an spatial proximity Yet another hypothesis is that the observe suppression is simply the result of the spatial proximity an alignment of the preferre an nonpreferre olor pathes an that the ontour geometry at the T-juntion itself oes not play a role. In our stimulus esign, all aiental ontour features are forme at a triolor juntion with a ontinuous oluing ontour forming the hat of the T-juntion. To etermine whether a triolor juntion alone, without the ontinuous oluing ontour, was suffiient to suppress responses to preferre sharp onvexities, we stuie responses to stimuli that were equivalent to the ombination stimuli being viewe through a irular aperture (Fig. 2h). In these stimuli, although the triolor juntion is preserve, the oluing ontour is trunate. Unlike ombination stimuli, triolor ontrol stimuli evoke responses that were omparable with primary shape responses (Fig. a ) an suppression was signifiantly less evient (S ol.64 vs S triolor.22). Of the 11 ells teste on the triolor ontrol, showe statistially signifiant ( p.5) lower suppression for the triolor ontrol stimuli (Fig. e). This suggests that a triolor juntion alone is not suffiient an that the ontinuous oluing ontour (the hat of the T-juntion) may be neessary for suppresse responses. To further aress this point, we quantifie the frational suppression of ontext stimulus responses (i.e., the foregroun or oluing stimulus responses) when presente in ombination with primary shapes. Context stimuli spatially ajoin the primary shapes but they lak a urvature isontinuity along their bounary. On 32 ells that showe more than half-maximum responses on at least two ontext stimuli, we alulate frational suppression of responses to ombination stimuli relative to ontextual stimuli. Responses to ontext stimuli also showe some suppression (average frational suppression of.15) when in ombination, but the suppression of ontextual responses was signifiantly less (ranomization paire t test, p.1) than suppression of primary shape responses. These results support the hypothesis that spatial proximity alone is not suffiient, an ontour geometry, whih itates pereption (Chapanis an MCleary, 1953), is likely to be important. Another simple hypothesis is that energy in a speifi spatial frequeny ban unerlies the observe suppression. All ombination stimuli have a harateristi preferre olor/nonpreferre olor/preferre olor pattern in the stimulus image beause of the onvex projetions in the ontextual stimuli. This reates energy in a speifi spatial frequeny ban, an neurons tune to these spatial frequenies oul provie the inhibitory rive to ause the suppression. Suh a mehanism will preit iential levels of suppression for preferre ontours an their 18 rotate images (attributable to the symmetry of the Fourier transform of real-value images), but our ata emonstrate that this is not the ase. For example, ompare levels of suppression for stimuli at 27 an 9 in Figure 3.

10 Bushnell et al. Partial Olusion an Shape Enoing in V4 J. Neurosi., Marh 16, 11 31(11): a Primary shapes b Context Stimuli Combination =.1 RFra e =.2 RFra g Sol Aross all ells, frational suppression base on preferre stimuli was signifiantly greater than that base on 18 rotations (average frational suppression aross ells for preferre shapes,.32; for 18 rotate shapes,.5; ranomization t test, p.1). This is espite the fat that suppression for 18 rotate stimuli is biase towar larger values as a result of the low firing rates. Lateny of suppression onset Suppresse enoing of aiental ontours emerge soon after response onset. In the example neuron illustrate in Figure 11a, onset of suppression is very early, an the timing is omparable with when shape-seletive responses appeare, i.e., ombination responses (Fig. 11, blue) an nonpreferre primary responses (Fig. 11, green) have a similar temporal profile. For this neuron, lateny of response onset was 42 ms for preferre primary shapes (Fig. 11a,, re) an for ombination shapes (Fig. 11b,, blue). The time of suppression onset, whih was evaluate as the time at whih the ifferene between the primary responses f =.5 RFra Spatial separation ( RF raius) Figure9. Spatialseparationresults: exampleanpopulation. a f, Exampleell. AllonventionsasinFigure3. a, Primaryshape responses. This neuron respone preferentially to shapes with a sharp onvexity ajoine by a onavity at the bottom. b, Responses to ontextual stimuli were uniformly weak., Combination responses reflet suppression of preferre primary shape responses uner partial olusion ontext. f, A small spatial separation (,.1 RF raius; e,.2 RF raius; f,.5 RF raius) between primary an ontextual stimuli ramatially erease suppression. g, Population results. Spatial separation (x-axis) versus frational suppression(y-axis). Most neurons(14 of 17; see Results) show a negative orrelation between frational suppression an spatial separation. Note for the interrupte x-axis: for one neuron, the maximum separation teste was 1.25 RF raius. spikes/se an ombination responses reahe statistial signifiane (see Materials an Methos), was 47 ms. In other wors, suppression emerge 5 ms after the onset of preferre primary responses. Aross all 76 neurons that showe statistially signifiant suppression, mean lateny for response onset to preferre primary shapes was 55 ms (SD 25 ms); mean onset lateny for ombination stimuli was longer (63 ms) an more variable (SD 49.5 ms). Suppression of ombination responses emerge, on average, 13 ms (SD 25 ms) after the onset of preferre primary responses. To failitate iret omparison, following Zhou et al. (), we quantifie the ifferene histogram (Fig. 11,e, blak) between preferre primary responses an the orresponing ombination responses; the time to halfmaximum of the ifferene histogram from stimulus onset was measure as the lateny of suppression onset. By this metho, lateny of suppression onset was 46 ms from stimulus onset for the example in Figure 11a an 63 ms for the population (Fig. 11e). This is earlier than the emergene of borer ownership signals in V2 ( 68 ms) quantifie by the iential metho (Zhou et al., ). Disussion We stuie the responses of V4 neurons that were seletive for ontour urvature to shapes presente in isolation an in the presene of ajoining ontext to isover whih ontours of partially olue objets are enoe. Our results iniate that V4 responses that enoe sharp onvexities at the T-juntion between objets are suppresse by the presene of the ontextual stimuli, whereas those enoing broa onvex urvatures are essentially unaffete. Hypotheses base on loal olor, luminane ontrasts, an response normalization mehanisms o not explain these results. Control experiments suggest that spatial proximity of the ontextual stimulus alone is not suffiient to ause the suppression; the bouning ontours of the primary an ontextual stimuli nee to form a T-juntion. Our finings parallel behavioral results that emonstrate longer reation times when shape mathing is base on sharp onvex an onave features at the juntion between shapes ompare with when those features boun a shape in isolation (Rensink an Enns, 1998). These results are also onsistent with shape-theoreti an psyhophysial results iniating that the sharp onvexities at the juntion between shapes are pereive as aiental ontour features (Helmholtz, 199; Chapanis an MCleary, 1953). The ontinuous broa onvex urvatures, onversely, are statistially more likely to be real ontours even in the presene of ajoining stimuli. Below, we isuss the impliations of our finings to the representation of partially olue objets in

11 22 J. Neurosi., Marh 16, 11 31(11):12 24 Bushnell et al. Partial Olusion an Shape Enoing in V4 V4, the generation of borer ownership signals, an the segmentation of images along the ventral stream. In natural senes, when objets are partially olue, the visual system must segment the sene an assign bounaries to the appropriate objets. Shape theorists sine the time of Helmholtz (199) have postulate that T-juntions serve as the primary ue for olusion (Guzman, 1968; Clowes, 1971; Huffman, 1971; Waltz, 1975). Psyhophysial finings also suggest that T- juntions are important for pereption uner partial olusion (Eler an Zuker, 1998; Rubin, 1). In pitorial isplays (without epth ues), it is only at T-juntions that the epth orering of overlapping objets an be etermine (Fig. 12a), an, base on this information, the intervening ontour an be assigne to the appropriate objet (i.e., enoing of borer ownership) (Nakayama et al., 1995). Although borer ownership signals have been oumente in the primate brain (Zhou et al., ), it is not known how information at T-juntions leas to their generation. Moels of image segmentation an borer ownership have typially invoke the expliit or impliit enoing of T-juntions, followe by instantiation of rules to ientify the iretion of the oluing bounary at the T-juntions (Fig. 12a) (Saja an Finkel, 1995; Zhaoping, 5; Craft et al., 7). Our results provie physiologial eviene for how information at T-juntions oul influene ontour enoing an raises the possibility that the importane of T-juntions oul ome from the suppression of aiental ontours rather than their expliit representation (whih has not been physiologially emonstrate). For a omplex visual sene with partially olue objets, suppresse enoing of sharp onvexities an onavities at the juntion between objets woul result in a representation that more strongly enoes the non-aiental real ontours of the various objets (Fig. 12). Suh a representation woul be equivalent to ientifying the iretion of the oluing bounary at the T-juntion beause, one aiental ontours are suppresse, the only ontour enoe at the T-juntion will be the oluing ontour. Then, previously propose ollinear an o-irular failitation mehanisms oul lea to appropriate borer assignment of the ontour away from the T-juntions (Saja an Finkel, 1995; Zhaoping, 5; Craft et al., 7). Preferential enoing of non-aiental ontours oul also provie an ieal population oe for reognition by failitating the bining together of only those features that are atually assoiate with an objet. Our results a Primary shapes b Combination Tri-olor ontrol spikes/se Spikes/se Combination responses Tri-olor responses Primary response (spikes/s) S ol.6 e.4.2 Combination stimuli Tri-olor ontrol Figure. Results from an example neuron on the triolor ontrol experiment. All onventions as in Figure 3. a, Primary shape responses. This neuron respone preferentially to primary shapes in the orientations. b, Combination responses reveale strong suppression of preferre primary shape responses (S ol.64) uner partial olusion ontext., Triolor ontrol stimuli are visually equivalent to the ombination stimuli viewe through a irular aperture. Responses were omparable with primary shape responses (S triolor.22). Speifially, responses were not as suppresse as in b., Satter plot of primary shape responses(x-axis) versusombinationantriolorresponses(y-axis). Triolorresponseswerebetterorrelatewithprimaryshape responses than ombination responses. e, Population results for all 11 neurons teste on the triolor ontrol. Eah line represents a ell. y-axis plots frational suppression base on ombination (left) an triolor stimuli (right). Preferre primary Combination Non-preferre primary a b spikes/se Normalize response Time from stimulus onset (ms) Population Time from stim. onset (ms) e Preferre primary Combination Non-preferre primary Pref. primary-combination Figure11. Latenyofsuppressiononset: exampleneuronanpopulationresults. a, Rasterplotsshowsingle-trialresponses foranexampleneuronforpreferreprimaryshapes(stimulithatevokegreaterthanhalfmaximumresponses)(a), orresponing ombinationstimuli(b), annonpreferreprimaryshapes(stimulithatevokelessthanorequaltohalf-maximumresponses)()., Peristimulus time histograms for preferre primary shapes, ombination stimuli, an nonpreferre primary shapes were smoothe with a Gaussian 5 ms. Differene between preferre an ombination histograms is shown in blak. Time relative to stimulus onset is plotte along the x-axis. Gray lines represent SEM. e, Population histograms base on normalize responses of 76 of 129 neurons that showe statistially signifiant suppression of ombination responses.