Motion is a potent cue for breaking camouflage in the natural world. To understand the neural basis of this phenomenon, one must utilize moving shapes defined by coherent motion of random texture elements against a similar, but stationary texture. To investigate how well neurons in area V4 process this novel, ecologically relevant stimulus and to compare shape selectivity for these shapes with static and other moving shapes, we tested V4 neurons with 5 static or moving shapes defined either by luminance or kinetic cues. The kinetic cues included a temporal frequency cue due to the difference in temporal frequencies of the moving dots inside the shape boundary and stationary dots outside the boundary. Therefore, static opponent motion–defined shapes without this cue were tested as an additional control. Approximately 44% (95/216) of V4 neurons showed shape selectivity. Analyses of these selective neurons both at single-neuron and population levels revealed that the shape-selective V4 neurons responded selectively to the moving kinetic shapes and that these neurons demonstrated robust invariance for shape preference across different shape conditions. Cue-invariant shape selectivity was more pronounced when kinetic shapes included the temporal frequency cue. This invariance may be rooted in nonlinearities occurring early in the visual pathway.
Motion is a potent segmentation cue that can defeat camouflage in the natural world. For instance, an animal with an appearance similar to that of its surroundings will be invisible as long as it remains still. As soon as it moves, the motion breaks the camouflage and the animal becomes visible. It is not only important to detect the motion but also to be able to categorize what is moving, for example, a predator or prey. Thus, it is likely that the ventral visual stream, which is involved in coding visual information that is relevant for object categorization and recognition, responds selectively to shapes that are defined by camouflage-breaking motion.
A visual stimulus that closely mimics this breaking of camouflage is a shape made of random texture elements that move coherently against a static pattern of the same texture elements. Hence, the shape is defined by its motion on the static background. Stimuli similar to these moving kinetic shapes have been used to study the orientation and direction selectivity of dorsal stream areas V3A (Zeki et al. 2003) and MT/V5 (Albright 1992) and in early visual areas (Chaudhuri and Albright 1997; Leventhal et al. 1998; Bourne et al. 2002; Lui et al. 2005). Because such stimuli have never been used to probe the responses of neurons in the ventral visual system, it is not known whether neurons in these visual areas respond selectively to such moving kinetic shapes. Although several studies have shown that ventral stream areas are indeed involved in processing motion-defined contours (Sary et al. 1993, 1995; Marcar et al. 1995, 2000; Mysore et al. 2006), these studies employed static shapes that were defined by opponent motion of the texture elements inside and outside of the shapes. Thus, given the ecological importance of the moving kinetic shapes and the known role of ventral stream structures in extracting shape information, we measured the shape selectivity of single neurons in area V4, an intermediate ventral visual stream area, to moving kinetic shapes.
The moving kinetic stimulus differs from the opponent motion kinetic stimulus that was used in previous studies (Sary et al. 1993, 1995; Marcar et al. 1995, 2000; Mysore et al. 2006) in 2 important ways. Firstly, the boundaries of the opponent motion kinetic shapes were static, which is somewhat artificial compared with the more natural camouflage-breaking phenomenon in which the shape boundaries are moving. Secondly, in order to detect the opponent motion kinetic shapes, the system needs direction-selective mechanisms, whereas for the moving kinetic shapes, a mechanism that detects differences in temporal frequency is sufficient. This cue is a consequence of the differences in temporal frequencies between coherently moving random dots inside the moving shape boundary and stationary dots outside of the boundary. Given the significance of this temporal frequency cue in boundary segmentation (Sary et al. 1994; Nawrot et al. 1996), we predicted that the selectivity of single V4 neurons for such moving kinetic shapes should at least approach that of luminance-defined shapes.
It was with this hypothesis in mind that we chose to compare the responses of single V4 neurons with moving kinetic and moving luminance-defined shapes. In addition, we measured the responses to static kinetic shapes and static, luminance-defined shapes in order to assess the contribution of the motion per se to the responses. Indeed, a comparison of the responses to moving and static shapes is, by itself, of interest. It has been established that as long as speeds are sufficiently slow, human acuity remains high (Mckee et al. 1986), enabling fine shape processing. Yet, very few studies have investigated the neuronal underpinnings of shape processing for moving stimuli. Indeed, the classical paradigm for exploring shape processing in the ventral pathway has been the static presentation of shape stimuli that are flashed on and off to an awake animal (e.g. Schwartz et al. 1983; Sary et al. 1993; Logothetis and Sheinberg 1996; Pasupathy and Connor 2001; Tamura and Tanaka 2001). A notable exception is the study in the anesthetized monkeys (Kobatake and Tanaka 1994) in which slowly moving stimuli were presented, but even here no comparison between shape selectivity for static and moving stimuli was attempted.
Therefore, in order to study the response properties of dorsal V4 neurons to a novel, but more natural, camouflage-breaking shape stimulus and to address whether the shape preferences are similar across different cues, for example, luminance or kinetic shapes, we presented shape stimuli designed according to a 3-way factorial design. Five different shapes presented on a static background could be defined by a luminance or motion difference and could be either static or moving. To ensure generality, all motion in the stimuli was tested in 2 directions, and to compare with previous results, static shapes defined by opponent motion were also included in the testing.
Subject and Surgery
We recorded single neurons in dorsal area V4 from 2 awake, behaving monkeys (Macaca mulatta). The head post and the recording chamber were implanted using isoflurane anesthesia under sterile conditions. We utilized anatomical magnetic resonance imaging scans to localize area V4 and to position the recording chamber. All experimental procedures were in accordance with the national and European guidelines and were approved by the K.U. Leuven ethical committee.
Subjects performed a passive fixation task in order to obtain a juice reward. Eye positions of the subjects were tracked using an infrared cornea reflection method (ISCAN). Subjects maintained a stable fixation on a red dot (fixation window size, 1 × 1 degrees for both subjects). We used glass-coated tungsten electrodes with an impedance of approximately 1 Mohm for the standard extracellular recording details of which are described elsewhere (Vogels and Orban 1990; Schoups et al. 2001; Mysore et al. 2006).
Stimuli and Testing
All the shape stimuli were stored as 640 × 480–pixel image sequences and were presented as a continuous “movie” of frame sequences on a CRT monitor at a frame rate of 60 Hz, while the subjects viewed the screen from a standard distance of 75 cm (37.2-degree diagonal; width and height 29 and 22 degrees, respectively).
We used 5 shapes, which were a subset of the 8 shapes used in the previous V4 (Mysore et al. 2006) and IT (Sary et al. 1993) studies, to test shape selectivity of these neurons. These 5 shapes were selected on the basis of a cluster analysis of the population response to the set of eight 2-dimensional (2D) shapes tested in the Mysore et al. (2006) study. We eliminated the 3 shapes that clustered most closely with other shapes. Shape stimuli were presented on a random texture background (50% white and 50% black pixels, pixel size: 2.9 × 2.9 min.arc, background area: 664 degrees2). Each shape was available in 5 sizes ranging in area from 1.1 to 26.8 degrees2. All shapes of any given size were of equal area, and the size chosen for testing was adapted to the size of the receptive field (RF) of the neuron. The median area of the shape stimuli used in the current study was 5.3 degrees2 (first and third quartiles 2.3 and 5.3 degrees2).
Shape stimuli utilized in the factorial design were defined either by motion contrast (kinetic shapes) or by luminance contrast and were presented against a static random textured background.
Kinetic shapes were defined by coherent motion of the random textured pattern (50% black and 50% white pixels) inside the shape boundary at a standard speed of 3 degrees/s against a similar (50% black and 50% white pixels) yet static textured background. Coherently moving dots had unlimited lifetimes and appeared and disappeared at the shape boundaries. The shape boundaries themselves remained either static (kinetic static; KS) or traversed the RF of the sampled neuron at 3 degrees/s (kinetic moving; KM) against the static texture pattern, resulting in 2 kinetic shape conditions (KS, KM; Fig. 1A). The KS shape condition, in turn, consisted of 2 subconditions: one where the pixels inside the static shape boundary moved coherently (KSH; horizontal) to the right and one where the pixels moved upward (KSV; vertical) (Fig. 1A bottom left panel). The same 2 shape subconditions were utilized for the KM condition (Fig. 1A bottom right panel), except that here the shape boundaries moved through the RF of the neuron at the same speed (3 degrees/s) and direction, either to the right (KMH; horizontal) or upward (KMV; vertical), as the dots moving inside the shape. These 2 kinds of motion-defined shapes (KS and KM) are visible against a static background by virtue of 3 distinct motion-related cues: a temporal frequency difference between the coherently moving dots inside the shape boundary and the static random texture outside of the boundary, a velocity difference between the dots inside and outside the shape boundary, and a dynamic occlusion cue at the shape boundaries due to the appearance and disappearance of dots at the boundaries that are not parallel to the motion direction (Sary et al. 1993; Sary et al. 1994; Nawrot et al. 1996; Vogels and Orban 1996). Thus, the detection of these kinds of motion-defined shapes does not necessarily require direction-specific velocity mechanisms.
As with the kinetic shapes, 2 shape conditions were generated using luminance-defined shapes: one where the shapes boundaries remained static (luminance static, LS) and another where the shapes traversed through the RF of a neuron at the standard speed of 3 degrees/s (luminance moving, LM). The shapes in the LS condition had one of 2 contrast polarities defining 2 subconditions, black shapes (LSB), with 20% white pixels, or white shapes (LSW) with 20% black pixels. Luminance of the bright pixels measured 39.5 cd/m2 with a Michelson contrast close to 100%. The shapes were presented against a static background with 50% black and 50% white pixels (Fig. 1A, top left panel). The 20% of opposite-polarity pixels was used to reduce the saliency of the luminance cue defining the shape because the kinetic cue is weaker than the luminance cue. In the LM condition (Fig. 1A, top right panel), 4 shape subconditions were generated by the combination of the 2 polarities and 2 directions of motion: black and white luminance-defined shapes traversed the RF in the horizontal direction (LMBH and LMWH; to the right) or in the vertical direction (LMBV and LMWV; upward).
These kinetic and luminance-contrast stimuli (LS, KS, LM, and KM) defined a 2 × 2 × 5, 3-way factorial design with the cue defining the shape (luminance versus kinetic) as the first factor, the nature of boundary of the shape (static versus moving) as the second factor, and the shape itself (5 simple geometric shapes, see Fig. 1A) as the third factor.
We also tested the neurons with static shapes defined by relative motion (Sary et al. 1993; Vogels and Orban 1996; Mysore et al. 2006), which were not a part of the factorial design. These opponent motion kinetic shapes (omKS) remained static (Fig. 1B) but were defined by relative motion where the pixels inside and outside (background; 664 degrees2) the static boundaries moved in opposite directions at equal speeds (3 degrees/s). They contained 2 motion-related cues: a difference in the direction of motion of dots inside and outside the shape boundary and the dynamic occlusion cue due to the appearance and disappearance of dots at the static shape boundaries (Sary et al. 1993, 1994). Note that the latter cue renders only those contours visible that are not parallel to the motion directions, implying that the full shape is defined by the direction difference only. Here too, both horizontal (omKSH) and vertical motion (omKSV) subconditions were tested. These omKS shapes were similar to the kinetic shapes used in the previous IT study (Sary et al. 1993) and exactly the same as in our previous V4 (Mysore et al. 2006) study.
The shape test included 66 stimuli: 50 shape stimuli of the factorial design, 10 omKS stimuli, and 6 other control stimuli that are not relevant to the present report. The 50 shape stimuli of the factorial design consisted of the 2 luminance (5 shapes × LSB and 5 × LSW) and the 2 kinetic subconditions (5 × KSH and 5 × KSV) for the static shapes, as well as the 4 luminance subconditions (5 × LMBH, 5 × LMBV, 5 × LMWH, and 5 × LMWV) and 2 kinetic subconditions (5 × KMH and 5 × KMV) for the moving shapes. Additional 10 stimuli were generated using omKS shapes (5 × omKSH and 5 × omKSV). Notice that the “shape test” included 12 subconditions out of which 10 subconditions constituted the factorial design.
After 1000 ms of stable fixation during which subjects viewed a red spot (0.2 degrees) against the static random dot background, the shape or control stimuli were presented. The static shapes and the control stimuli were presented for 500 ms, but the motion amplitude, and hence duration, in the moving shape conditions depended on the size of the RF of the neuron because the moving shapes needed to completely traverse the RF of the neuron. The movement encompassed a distance of 9 degrees (3000-ms stimulus duration at 3 degrees/s speed) in the vast majority (182/216) of the sampled neurons. In the remaining cells (34/216), the stimulus duration for the moving shapes was 4000 ms (corresponding to motion amplitude of 12 degrees). The static random dot background presented before stimulus onset was also shown for 300 ms as a poststimulus period in all conditions. Across trials, the static shape, the moving shape, and control stimuli were interleaved. In the current study, neurons were tested with at least 5 presentations of each stimulus of the shape test. The median number of presentations was 8; first and third quartiles were 8 and 10, respectively.
We adopted the following standardized testing procedure. After isolating a single neuron, we flashed small checkerboards or luminance gratings (1- to 2-degree squares) in a regular pattern across the lower visual quadrant (range of 0–14 degrees) to plot the RF of the neuron (search test). The estimated center of the RF was defined as the position where the flashed stimulus elicited the maximum response as judged from inspection of the peristimulus time histograms (PSTHs) of the responses for the different (n = 48) positions. The size of the shape stimulus was adjusted for each neuron so that the RF encompassed the entire shape (median ratio of shape size to RF size was 0.57). Median RF size (square root of RF area) was 3.12 degrees (quartiles 2.73–7.14 degrees). The size and center of the RF, estimated online during the experiment, corresponded well with those of the off-line analysis of the responses from the search test. Once the size and approximate center of the RF had been determined, we tested the neuron with the shape test. Testing every isolated single neuron by including all conditions of interest in a single test enabled us to explore the selectivity of V4 neurons for static and moving shapes defined by luminance contrast and by motion contrast in an unbiased manner.
Analysis of Neuronal Activity
Analysis Window for Static and Moving Shapes
Because the moving shapes, unlike static shapes, traversed the RF completely in either the horizontal or vertical direction, stimulus durations were different for static (LS and KS) and moving shape boundary (LM and KM) conditions. Hence, we set the analysis windows differently when computing spike activity elicited by these 2 shape conditions (LM and KM). For static shapes, we compared baseline average firing rates 500 ms before the stimulus onset with the average firing rates after the stimulus onset computed in a 500-ms window from 50 to 550 ms after stimulus onset. For the moving shapes, first we pooled firing rates for all horizontal motion conditions and for all vertical motion conditions separately using 20-ms bins. We then calculated mean and standard deviation of the firing rates during the baseline period (500 ms before the stimulus onset). A “cutoff” firing rate was determined by taking this mean firing rate plus 2 standard deviations. Finally, the analysis window (“cutoff window”) was defined as the part of the stimulus period during which the firing rate exceeded the “cutoff” rate. This was done separately for horizontal and vertical conditions, resulting in different analysis windows for each of these 2 directions.
Because the moving shapes traversed through the RF, the neuron may respond to parts of a shape stimulus as that stimulus enters and leaves the RF. Hence, the cutoff window, as defined above may also include responses to the leading and the trailing edges of shapes. In an additional control analysis, we utilized only that part of the response window for which the shape was estimated to lie entirely inside the RF. For this, we calculated the time window (“restricted window”) for which the shape would lie completely inside the RF, computed separately for the 2 orthogonal directions. The duration of this restricted window was computed using the shape having the maximal width and height. All analyses reported in the results were carried out with responses calculated using the cutoff window except for the above mentioned control analysis.
Analysis for Responses to Shape
A split-plot analysis of variance (ANOVA) (within-trial factor: baseline response; between-trial factor: shape) comparing firing rates during the baseline period and the firing rates during the response window was carried out (main effect of the within-trial factor) to establish whether a neuron responded significantly (P < 0.05) during the stimulus periods. We applied this split-plot ANOVA separately for each of the shape subconditions. Selectivity to shapes was determined by using a 1-way ANOVA (P < 0.05) on the net responses (after subtracting the baseline activity), again performed separately for each of the shape subconditions.
Analysis of the Factorial Design
In keeping with the factorial design, we utilized a 3-way factorial (2 × 2 × 5) ANOVA to examine the main effect of shape and to look for interactions between the factors (2-way interactions between cue [luminance/kinetic] and shape or between the nature of the boundary [static/moving] and shape and a 3-way interaction between cue, nature of the boundary, and shape). Notice that the number of stimuli in the 4 shape conditions is unequal because the LS, KS, and KM conditions included 2 shape subconditions (e.g., LSB and LSW or KSH and KSV), and hence 10 stimuli, but the LM condition, had 4 subconditions (LMBH, LMBV, LMWH, and LMWV) and thus 20 stimuli. Nonetheless, we were able to analyze the data with a symmetrical design because we observed high correlations between responses to the subconditions within each of the shape conditions. For the LS shape condition, we observed high correlations between responses across shapes (n = 5 shapes) to the 2 contrast polarities (subconditions), with a median correlation between LSB and LSW of 0.80 (first and third quartiles, 0.36 and 0.90). High levels of correlation were also observed between the responses to 2 directions of the KS condition, with a median correlation between KSH and KSV of 0.85 (0.49 and 0.95). The same was true even for the moving shape conditions. For the LM condition, the median correlation between LMBH and LMBV was 0.74 (0.38 and 0.90) and that between LMWH and LMWV was 0.74 (0.43 and 0.86). For the KM condition, median correlation between responses to KMH and KMV was 0.77 (0.34 and 0.89). The distributions of the correlation coefficients were all positively skewed and were significantly different from zero (t-test, single sample, P < 0.05, for the 2 static and each of the 3 moving shape condition correlations). Given the high levels of correlations between subconditions, we will report data that used the most responsive subcondition of a given shape condition, defined as that with the maximum average response across all 5 shapes of a given subcondition, as representative of that shape condition (e.g., choosing either LSB or LSW for LS condition), which enabled us to use a symmetrical design for the analysis. However, in an additional control analysis, we employed the least responsive subconditions instead of the most responsive subconditions. Factorial ANOVA was performed on the net responses of the selected subconditions.
For each neuron, we assessed the separability of the stimulus dimension shape from the combination of the 2 other factors defining the shape conditions, the cue, and the nature of the boundary, using an estimation of the tuning along each of the 2 dimensions obtained by summing over the other dimension. Like the ANOVA, this separability measure uses the representative subcondition for each of the 4 shape conditions and is calculated by estimating the predicted response and correlating this estimation with the observed responses. First, the gross (baseline not subtracted) responses of each neuron for each representative shape subcondition is summed across the 5 shapes, producing an estimate of the variation of the responses with the 4 shape conditions (LS, LM, KS, and KM) across shapes. The responses to each shape were also summed across all 4 shape conditions, producing an estimate of the variance of the responses with the 5 shapes. The 4 × 5 (4 shape conditions × 5 shapes) 2D matrix product of these 2 marginal sums is the predicted response of the neuron to a particular shape in a particular shape condition (LS, LM, KS, and KM), assuming separability of the stimulus dimensions shape and shape condition. The degree of separability of the 2 stimulus dimensions was defined quantitatively as the squared Pearson correlation (r2) between the observed and the predicted responses for each cell (Mazer et al. 2002; Kayaert et al. 2005). We performed 1000 random permutations of the observed responses, calculated the predicted responses with each of the permuted responses, and correlated the permuted responses with the predicted responses. We considered the separability measure, that is, the r2 for the observed and predicted responses, for a particular neuron to be significant if the observed correlation exceeded the 95th percentile of the frequency distribution of the correlations for the permuted data.
Analysis of Eye Positions
Our shape stimulus set included moving shapes that could be a potential source of pursuit eye movements for the subjects. Therefore, in addition to sampling neuronal activity, we also sampled and saved the eye position of subjects in about one-third of the recordings of Monkey I (60/156) and in all 60 neurons sampled in Monkey V in order to ascertain whether direction of eye movements of either subject correlated in any way with the direction of motion in the stimuli.
We used eye position data during the entire stimulus presentation and calculated the standard deviation of horizontal (x) and vertical (y) eye positions (oversampled at 200 Hz) within a given trial. Stimulus duration was 500 ms for all static stimuli; however, stimulus duration varied for moving stimuli (see above), depending on the size of the RF of the neuron.
In order to examine whether there was a correlation between the direction of the subject's eye movements and the direction of motion of the shape stimuli, we first calculated the mean eye positions in both the x and y directions during the baseline period. Then, the mean x and y positions were similarly derived for the entire stimulus presentation period. From these 2 means, we then computed an eye movement vector for each trial. The origin of the vector in each trial was the mean x and y positions obtained from the baseline period, and the endpoint of the vector was the mean x and y positions during the stimulus presentation (Mysore et al. 2006).
As a first analysis, we computed the circular correlation (Batschelet 1981) between direction of the stimulus and the eye movement direction. In computing a given correlation, we used all trials and all shapes belonging to a particular direction for the analysis. Correlations were calculated for 8 shape subconditions of the KS, KM, and LM conditions in both animals.
An additional analysis was performed on each of the 4 pairs of these 8 subconditions. We compared the distributions of the eye positions during the stimulus periods for orthogonal (horizontal and vertical) directions by using a 2D Kolmogorov–Smirnov test (Williams et al. 2003; Mysore et al. 2006). For this analysis, we pooled data from all trials and from all neurons (n = 120) corresponding to horizontal and vertical conditions.
In addition, we computed the eye velocities in the different shape conditions. First, the eye positions were resampled at the sampling frequency of the ISCAN (60 Hz), and the eye velocities in x and in y dimensions were calculated by differentiating the x and y eye position traces. This differential represented the instantaneous velocities in x and in y directions expressed in degrees per second. From these eye velocity traces, we calculated the amplitudes of the eye velocity vectors by taking the vector sum of the velocities in x and y dimensions. We identified microsaccades as eye movements for which the speed exceeded 8 degrees/s. The x and y segments of the velocity data that corresponded to the microsaccades were removed from the velocity traces. The missing segments of such velocity trace both in x and y dimensions were filled in with linear interpolations of the velocities before and after the microsaccades.
To assess whether the direction of motion of the shape stimuli systematically altered the velocity of the subject's eye movements, we analyzed the eye velocity data obtained as explained above separately for horizontal (x) and vertical (y) dimensions. If the subjects followed the stimulus direction, one would expect a change in the velocity profile of the subject's eye movements. Hence, we analyzed the velocity data for the 2 directions of motion of the shape stimulus using factorial ANOVAs. First, we calculated a mean velocity in horizontal (x) and vertical (y) dimensions, by averaging the velocities within the stimulus period of each trial. We performed factorial ANOVAs on the mean velocity data for static (KS) and moving (LM and KM) shape conditions separately. For the KS condition, we performed 2-way ANOVAs with the 5 shapes and the direction of stimulus motion (KSH or KSV) as factors, whereas we performed 3-way ANOVAs for the moving shape conditions with the cue (LM or KM), direction of stimulus movement (horizontal or vertical), and the 5 tested shapes as factors. As with the KS shape condition, we also analyzed eye velocity data for the omKS shapes again performing 2-way ANOVA with 5 shapes and 2 directions of motion (omKSH and omKSV) as factors.
Complete data were collected from 216 single V4 neurons in 2 awake, behaving macaques (156 from Monkey I, 60 from Monkey V). We sampled from both hemispheres in Monkey I (108 from the left hemisphere and 48 neurons from the right hemisphere), and eye position data were collected simultaneously along with single neuronal data for 60 neurons. In Monkey V, eye position data were sampled for all 60 of the neurons recorded from the left hemisphere of this animal. All neurons sampled were responsive (split-plot ANOVA; P < 0.05) to the shapes of at least one of the 10 shape subconditions of the factorial analysis tested. RF centers in our sample were located in the lower visual quadrant with a median eccentricity of 4.6 degrees (first and third quartiles 3.6 and 5.8 degrees, respectively). Histological verifications in Monkey V confirmed that the recordings were in dorsal V4. Monkey I is still participating in other experiments; however, anatomical MR images taken after the recording in Monkey I had been completed, using reference glass tubes filled with copper sulphate solution, verified that the recordings were in the prelunate gyrus. Furthermore, RF size as well as the topography of the sampled neurons matched with the reported results in dorsal V4 (Desimone and Schein 1987; Gattass et al. 1988).
Out of the 216 responsive neurons sampled, 82 (38%) neurons showed a significant (3-way ANOVA; P < 0.05) main effect of shape, 78 (36%) showed a main effect of cue, and 84 (39%) a main effect of the nature of the boundary. Among the 82 cells with a main effect of shape, 17 showed no interaction, 27 an interaction between shape and cue, 25 an interaction between shape and nature of the boundary, and 13 a 3-way interaction between all factors. Only 5 neurons showed no main effect of shape but an interaction between shape and one or two of the remaining factors.
We considered a neuron as being shape selective when it responded selectively (1-way ANOVA P < 0.005; Bonferroni correction for multiple comparisons [10 subconditions]) to shapes in at least one out of the 10 subconditions tested. In our sample of 216 neurons, 95 were selective for the tested shapes in at least one subcondition. The distribution of the preferred shapes of these 95 neurons is shown in Supplementary Figure 1 for each of the 4 shape conditions. These 95 shape-selective neurons were subjected to a range of analyses both at the single neuron and at the population level. One such shape-selective neuron is shown in Figure 2. This neuron displayed a main effect of shape and an interaction between shape and nature of the boundary. Owing to the different durations of static and moving stimuli, the responses to these 2 types are plotted separately: LS and KS conditions in upper part (A) and LM and KM conditions in lower part (B) of the figure. PSTHs clearly indicate that the neuron was selective for shape in each of the 4 static subconditions (LSB, LSW, KSH, and KSV), as confirmed by 1-way ANOVAs on the net responses (P < 0.005 for each subcondition). This neuron responded well to the shape “inverted F” but was not responsive at all to the octagon. Note also that this neuron preserved its preference for the shape “inverted F” in the 4 static subconditions. The bottom panel (B) shows the responses of the same neuron to the same 5 shapes when the shape boundary was moving. Here too, it is apparent that this cell was selective for moving shapes in all 6 shape subconditions (LMBH, LMBV, LMWH, LMWV, KMH, and KMV) with significant 1-way ANOVAs (P < 0.005) for each of these 6 subconditions. It is important to note that this neuron retained its preference to the shape “inverted F” in these 6 moving subconditions and thus remained selective for this shape in all 10 shape subconditions, both static and moving, of the factorial design. This is shown distinctly in Figure 3 in which the average net responses of this same neuron are plotted for all 10 subconditions of the factorial design: the 4 static subconditions in (A) and the 6 moving subconditions in (B).
Cue-Invariant Shape Selectivity for Static and Moving Shapes: Population Analysis
To describe the cue-invariant shape selectivity among these 95 shape-selective neurons, we compared shape preferences in the 4 shape conditions of the factorial design using the shape rank approach as that previously utilized in IT (Sary et al. 1993) and in V4 studies (Mysore et al. 2006). For each neuron, we ranked the 5 shapes in descending order of their average net responses to the 5 LS shapes (such that the most responsive shape in the LS condition is assigned rank 1 and the least responsive shape received a rank of 5). The resulting shape rank order was then applied to the net responses for the KS, LM, and KM conditions of the same neuron (e.g., if the most responsive shape in LS condition was “octagon” and the least responsive shape was “Triangle”, “Octagon” was assigned rank 1 and “Triangle” was assigned rank 5 for all the 4 shape conditions). After ranking the shapes in all shape conditions using LS as reference for each neuron, the net responses were averaged for each rank across all the shape-selective neurons (n = 95).
Figure 4A shows the mean responses as a function of shape rank for the 4 shape conditions. The curves show monotonic decreases in responses for all 4 shape conditions with increasing shape rank. This decrease in response with increasing shape rank was extremely significant: P < 0.00005 (1-way ANOVA) for the KS, LM, and KM conditions, indicating that the average shape preferences of the selective neurons were similar in the 4 shape conditions. The shape rank curves shown in Figure 4A were plotted using responses to the LS condition as the reference. The results were similar even when we ranked shapes based on the most responsive shape condition, rather than ranking on the basis of the LS response. We calculated the most responsive shape condition by averaging the net responses to the 5 shapes within a shape condition. The most responsive condition was determined as the condition having the highest average response among the 4 shape conditions. As demonstrated in Figure 4B, even when we ranked based on the most responsive shape condition, the decrease in the responses with increasing shape ranks were significant for all 4 curves (P < 0.0005; 1-way ANOVA). This demonstrates that on average, shape preferences were very similar in the different shape conditions, despite the fact that many cells showed interactions in the 3-way ANOVA. However, even in cells showing significant interactions in the 3-way ANOVA, shape rank curves indicated that the average shape preference was maintained across conditions (Supplementary Fig. 2).
We analyzed the responses to the moving shapes by using the cutoff window where we considered the cell to be responsive whenever it exceeded the set criterion (see Methods). In a control analysis, we chose an alternative restricted window (see Methods; Supporting text) for which the shape was completely inside the RF of the neuron. The shape rank curves obtained using the average net responses that were calculated using this restricted window again resulted in similar shape preferences across shape conditions (Supporting text; Supplementary Fig. 3A,B).
In order to use a symmetrical design in the factorial ANOVA, we utilized only the most responsive subconditions (see Methods). To determine whether the invariance observed when we used the most responsive subcondition is preserved when using the least responsive alternative instead, we also computed the shape rank curves by using the least responsive subcondition. As shown in Supplementary Figure 3C, the rank curves for KS, LM, and KM followed the reference condition LS when using the least responsive subcondition, conveying that shape preferences were similar across shape conditions.
Whereas the shape-selective neurons in V4 demonstrated similar shape preferences across shape conditions at the population level, the nonselective group of neurons, as one might expect, failed to conserve this property. We randomly selected 50 neurons that were clearly not selective for the shapes tested in any of the shape subconditions (P > 0.20 1-way ANOVA) and plotted shape rank curves using LS as the reference (Supplementary Fig. 3D). As expected, the shape rank curves shown were quite flat (1-way ANOVA; nonsignificant [ns]) except, of course, for the reference curve LS. Although the neurons that were not shape selective failed to show a similar shape preference for static and moving shapes, they did respond to the moving shapes. The response of all non–shape-selective neurons (n = 121) to the most responsive moving shape (LM) correlated significantly with their response to the static (LS) counterparts (r = 0.69; P < 0.00001). In fact, the correlation between responses in the LM and KS conditions for non–shape-selective neurons did not differ significantly from that for the shape-selective group of neurons (r = 0.67; n = 95 neurons; P < 0.00001; difference between correlations for the 2 groups: ns).
Thus, the shape rank analysis (Fig. 4; Supplementary Figs 2 and 3) provides compelling evidence that dorsal V4 is capable of representing simple forms in a cue-invariant manner and that this invariance is preserved across changes in the cue defining a given shape, for both static and moving shapes. The shape-selective V4 neurons retain a shape preference for camouflage-breaking stimuli (KM) that is nearly equal to that of static kinetic or static and moving luminance-defined shapes.
Cell-by-Cell Analysis of the Shape-Selective Neurons
The ranking analysis averages across neurons and provides information on the invariance at the population level. To quantify the degree of cue invariance for individual neurons instead of the population, we computed the degree of separability of the shape and shape condition factor for each of the 95 shape-selective neurons. This analysis quantifies the degree to which the 2 factors, shapes, and shape conditions (combining cue and nature of the boundary) are separable for each neuron. Figure 5A shows the distribution of the separability index (see Methods) for the 95 shape-selective neurons. Neurons varied in their degree of separability (Fig. 5B–F; Supplementary Fig. 4). The average separability index was large in the population of shape-selective neurons, with a median (and quartiles) of 0.83 (0.70–0.90), and individual separability values were statistically significant (P < 0.05; permutation test—see Methods) in a large majority (68%) of the cases. Thus, most single V4 neurons retain their shape selectivity for luminance- and motion-defined shapes, and whether the shape boundary is static or moving. In order to reveal any differential effect between the 2 factors cue and boundary type, which were combined in the preceding analysis, we also computed the distributions of the separability indices for pairs of shape conditions. As indicated in Table 1, these distributions were exceedingly similar for the 6 pairings that could be derived from the 4 shape conditions.
|n = 95 for each pair|
|n = 95 for each pair|
Although average separability values were high within the shape-selective neurons, conveying that, in general, these shape-selective V4 neurons are strongly invariant in their shape preference; some cells showed less separability of shape and conditions. Examples of a few such cells that displayed nonsignificant separability values are illustrated in Figure 5B,C, and D. It is quite evident that these neurons were not strongly invariant for shapes across conditions, in sharp contrast to the example neuron in Figure 2, the average net responses of which is shown again in Figure 5E, and another example neuron in Figure 5F, both of which, incidentally, had very high separability values. However, as one can see, the neurons illustrated in Figure 5B,C were not selective for shapes in all the 4 shape conditions that might suggest a correlation between the degree of shape selectivity and separability.
Degree of Shape Selectivity Analysis
To assess the degree of shape selectivity in a shape condition, we computed the ANOVA-derived strength of association index ω2 (see Methods) within each shape condition. Table 2 lists the median (and first and third quartiles) values of all shape-selective neurons (n = 95) for the 4 shape conditions, using the representative subconditions of each condition. The medians of the ω2 distributions were very similar for all 4 shape conditions. The average values were all close to 0.4, and none of the 6 pairwise comparisons between conditions were significant (Wilcoxon matched-pairs test). Also, the distributions of ω2 were significantly different from zero for all shape conditions (P < 0.00001, t-test single sample). Thus, degree of shape selectivity was highly similar for different shape conditions, indicating that overall, shape-selective dorsal V4 neurons display a similar degree of shape selectivity, regardless of how the shapes are defined and whether the shape boundary itself is stationary or moving. As expected, the neurons with a higher degree of shape selectivity, as shown by high ω2 indices, tended to have higher separability indices as well. There was a significant, relatively large correlation (r = 0.68, 0.70, 0.67, and 0.66 for LS, KS, LM, and KM, respectively) between the degree of selectivity and the separability index, suggesting that neurons showing less invariance in their shape selectivity (as the example neuron of Figs 5B,C) did so because of their reduced shape selectivity, at least for the stimuli tested.
Thus, from these analyses, both at the population and single neuronal level, we can conclude that single V4 neurons can respond selectively to moving kinetic shapes (KM) and that the shape preference and degree of selectivity are comparable to those for moving luminance-defined shapes and for static kinetic and luminance-defined shapes.
Time Course of Responses for Static and Moving Shapes Defined by Luminance or Motion Cues
Figure 6A plots the population PSTHs of the 95 shape-selective neurons for preferred and nonpreferred static shapes (LS or KS). For both LS and KS conditions, shape selectivity emerged at the onset of response. Responses to the LS shapes had strong transients and began approximately 20 ms earlier than the responses to KS stimuli. Note that the start of the motion in the stimulus was defined as the onset of the first frame showing displacement of the random texture pattern; in other words, coherent movement was already present within the first frame of the stimulus sequence in the KS condition, and thus, the 20-ms response delay is not an artifact of needing a second frame to define the motion. The average sustained responses to LS and KS shapes are very similar.
As for static shapes, responses to preferred and nonpreferred moving shapes diverged soon after the response began, as shown in Figure 6B (horizontal motion direction) and Figure 6C (vertical motion direction), for both LM and KM conditions. Because the stimulus amplitude varied among neurons, population PSTHs are plotted as a function of the percentage of the maximal amplitude of a given neuron. Unlike static shapes, there were no apparent transient responses to moving shapes because the stimulus gradually entered the RF; we observed instead a gradual rise and fall in responses as the stimulus traversed the RF. Thus, any differences in time courses were not readily apparent between the LM and KM conditions. On the other hand, like sustained responses to static shapes, average responses to KM and LM were equally strong, again demonstrating that on average, shape-selective V4 neurons respond as well to the camouflage-breaking stimulus as they do to luminance-defined moving shapes.
Eye Movements Cannot Explain the Invariant Shape Selectivity
Figure 7A shows the average eye movement of the 2 monkeys in the horizontal (blue) and in vertical (red) directions for the preferred (most effective within a given subcondition) and least preferred shapes. The figure displays the average eye positions for the 8 subconditions involving motion in the stimulus, the static shapes defined by motion (KS), and the moving luminance-defined (LM) and kinetic (KM) shapes for the 120 cells for which eye position was sampled along with neural activity. For each neuron, raw eye position samples were aligned so that their mean position was at the center of the fixation window for each trial and then averaged across trials. It is apparent that the subjects fixated quite well, and eye position hardly changed during the trials. This was also the case for the eye positions measured for the example neuron in Figures 2 and 3 (Supplementary Fig. 5). Further evidence of the quality of fixation is provided by the small standard deviation of the eye positions within a trial. Median (first and third quartiles) values of these standard deviations across trials for all cells (n = 120) were 0.11 degrees (0.10 and 0.11) and 0.10 degrees (0.10 and 0.11) for the horizontal and vertical directions, respectively.
Eye movement analysis (see Methods) showed no correlation between the subject's eye movements and the direction of pixel motion in the motion-defined shapes or with the direction of motion of the shape boundaries. For static KS stimuli, only 4 out of 120 tests showed a significant (P < 0.05) circular correlation of motion direction and eye movement, no greater than expected by chance. The median (first and third quartiles) P values for the circular correlation coefficients in the static stimulus comparison equaled 0.71 (0.41–0.88) for the KS condition. For the moving shapes for which 3 correlation values were computed per neuron, only 17 tests out of 360 were significant in the analysis, again less than 5% of the tests. Here, the median (and first quartile–third quartile) P values were 0.74 (0.48–0.81), 0.47 (0.29–0.64), 0.64 (0.34–0.71) for LMB, LMW, and KM conditions, respectively. Additionally, none of the 4 comparisons between stimulus directions (1 static shape comparison between KSH and KSV, 3 moving shape comparisons between LMBH and LMBV, LMWH and LMWV, and KMH and KMV) reached significance (P < 0.05) in the 2D Kolmogorov–Smirnov analysis, implying very similar distributions for horizontal and vertical stimuli.
Analysis of eye velocities further showed that the monkeys did not make pursuit eye movements in the direction of stimulus movement. Average eye velocities in horizontal (x, blue) and vertical (y, red) dimensions are plotted in Figure 7B for the neurons (n = 120) for which we sampled eye movements. The velocity profiles were very similar for the horizontal and vertical dimensions with fewer than 5% microsaccades observed. We also analyzed the eye velocities more rigorously to check whether the stimulus direction had any effects on the eye velocity. Out of 240 ANOVAs performed (120 cells × 2 dimensions; see Methods) for the KS stimulus, only 3 showed any significant main effect (P < 0.05) or interaction between shapes and stimulus condition (KSH; KSV). The same was true for the moving shapes (LM and KM, 3-way ANOVAs; see Methods). Out of the 240 ANOVAs, only 3 showed any of the significant main effect (of factors cue, stimulus direction, or shapes) or interactions between the factors. The median P values of the ANOVA effects was 0.74.
Thus, rigorous analysis of eye movements clearly shows that none of the single-cell results we report here were influenced in any significant way by the subjects’ eye movement patterns.
Responses of Shape-Selective Neurons to Shapes Defined by Opponent Motion
Along with the shape stimuli that were included in the factorial design (LS, KS, LM, and KM), we also tested V4 neurons with static but opponent motion–defined shapes (omKS). Selectivity was reduced for the omKS shapes, compared with the 4 shape conditions included in the factorial design. To compare the degree of shape selectivity of V4 neurons with omKS condition and the shape conditions of the factorial design, we calculated the ω2 distributions for all the 216 tested neurons, separately for each shape condition. Table 3 lists the medians and quartiles of the ω2 distributions of the 5 shape conditions included in the shape test, using the representative subcondition for each condition. Although average ω2 was close to 0.2 in LS, KS, LM, and KM conditions, it reached only 0.12 in the omKS condition. Repeated-measures ANOVA for all 216 V4 neurons and the 5 shape conditions yielded a significant main effect of shape condition on the ω2 (P < 0.00005). Out of the 10 pairwise Bonferroni post hoc comparisons between shape conditions, the 4 tests comparing omKS with each of the 4 other conditions were significant, the 6 others were nonsignificant.
As a consequence of the reduced selectivity in the omKS condition, only 46 out of the 95 neurons that were shape selective in the 4 other conditions were also selective for omKS shapes as determined by 1-way ANOVA (P < 0.05). The example neuron shown in Figure 3 was one such neuron. Average net responses of this neuron to the 5 tested shapes is plotted in Figure 8A showing the representative subconditions of the 5 shape conditions. As one can see, the example neuron was also selective for omKS shapes (P < 0.002; 1-way ANOVA). More interestingly, even for shapes defined by opponent motion, this neuron retained its preference for the “inverted F” shape.
Among the 46 neurons that were also selective for static shapes defined by opponent motion (omKS), the shape preferences were similar in all the 5 shape conditions. This tendency is very conspicuous in Figure 8B that plots the shape rank curves for those 46 neurons, again using net responses to LS as the reference. Because we found a high correlation between responses to omKSH and omKSV (r = 0.72), we utilized only one of the 2 subconditions for the shape rank analysis, the most responsive subcondition as determined by average net responses.
The separability analysis further demonstrated that shape preference was less invariant in the omKS than in the other conditions. We calculated median separability between the 3 pairs of static shape conditions. Although the median (quartiles) value was 0.92 (0.89, 0.93) for the LS/KS pair, it reached only 0.85 (0.72, 0.89) for the LS/omKS pair and 0.81 (0.71, 0.89) for the KS/omKS pair. The percentage of cells with significant separability indices reached 71% for LS/KS pair, compared with 47% and 49% for the 2 other pairs. Repeated-measures ANOVA on the separability values of the 46 neurons for the 3 pairs yielded a significant main effect of pair (P < 0.00001). The Bonferroni post hoc comparisons between the 3 pairs yielded significant values for the 2 comparisons involving omKS but not for the third comparison between LS and KS.
In contrast to the overall population of 216 V4 neurons, there was no apparent difference in the degree of selectivity for omKS shapes compared with the other 4 conditions among the 46 neurons that were shape selective in all 5 shape conditions (0.42; 0.2–0.5 median and first–third quartiles). We found no significant differences when we compared ω2 values among these 46 shape-selective neurons across the 5 shape conditions (repeated-measures ANOVA; ns).
Interestingly, the average latency of the response onset and shape selectivity was longer for the omKS than for KS, the other motion-defined condition (significantly different from KS at ninth bin after stimulus onset (Fig. 8C). Moreover, the sustained part of the average response to the preferred shape in the omKS condition was lower than that in the KS and LS conditions (difference omKS − KS significant at P < 0.0001, Wilcoxon matched-pairs test). These differences suggest that the opponent motion cue is less effective for V4 neurons than the cues present in other, more natural, kinetic shape conditions.
In the omKS condition, the background encompasses the fixation point and moves in the opposite direction compared with the dots inside the shape boundary. This can potentially result in pursuit eye movements. To address this issue, we analyzed the eye movements in omKS condition in the same manner as we analyzed the eye movements for the other 4 shape conditions (see above). Supplementary Figure 6 shows the average eye positions and average eye velocities for all 120 neurons where eye movements were sampled during the omKS condition. The eye movement and eye velocity plots do not suggest any pursuit eye movements by the subjects. This was confirmed by 2-way ANOVAs of the average eye velocities in horizontal and vertical dimensions during each stimulus presentation for each cell. Out of 240 ANOVAs (120 × 2 dimensions), 4 were significant either for the main effect of shape or stimulus direction or showed an interaction between the 2 factors. The median P value of the ANOVA effects was 0.69. Thus, the shape selectivity in the omKS condition was not a consequence of stimulus-induced eye movements.
The present study is the first evidence showing that rather large proportions of dorsal V4 neurons can respond selectively to forms defined solely by their motion against a stationary background, that is, the type of camouflage-breaking stimulus most commonly encountered in a natural environment. Furthermore, we have shown here that the neurons selective for such stimuli are generally cue invariant with respect to their shape preference. This invariance was stronger for this type of camouflage-breaking stimuli than for the opponent motion stimuli used in previous studies (Sary et al. 1993, 1995; Marcar et al. 2000; Mysore et al. 2006).
Cue-Invariant Shape Selectivity in V4
The main finding of the present study is the remarkably robust invariance of the shape selectivity of V4 neurons. Shape selectivity remained very similar irrespective of the cue defining the shape, whether it consisted of luminance or motion and regardless of the nature of the boundary, whether it was static or moving. At the population level, shape preferences tended to be similar across shape conditions, suggesting that this sort of cue-invariant form processing is a general property of V4 neurons. This was confirmed by cell-by-cell analysis revealing that the majority of neurons had similar shape preferences, indicated by the large separability indices (median: 0.83). It is worth pointing out that the invariance applied equally well to the cue defining the shape as well as to the nature of the boundary.
Although shape selectivity exhibited a robust invariance, it is clear that the response strength nonetheless could differ between the conditions. This is illustrated by the example cell and has been observed in many of our previous studies (Sary et al. 1993; Vogels and Orban 1996; Mysore et al. 2006). The response of a single neuron provides ambiguous information because it can vary with both shape and shape condition (e.g., static versus moving shape or kinetic versus luminance-defined shape). For instance, even though the shape rank curves of Figure 4B demonstrate similar shape preferences across shape conditions, responses were stronger for shape rank 2 in the most responsive condition (yellow curve; C. Rank1) than for shape rank 1 in the second most responsive condition (C.Rank 2 green curve). Within each shape condition, however, the shape rank curves demonstrate that on average, the shape preferences themselves remained similar. Thus, the population of V4 neurons can signal shape across the different conditions by their relative activity.
There is ample evidence that the shape selectivity in V4 is only partial. It is most likely based on selectivity for subparts of an object rather than on global selectivity (Gallant et al. 1993, 1996; Kobatake and Tanaka 1994; Pasupathy and Connor 1999, 2001, 2002). Furthermore, when IT and V4 are directly compared using identical stimuli, shape selectivity is not as strong in V4 as in IT (Sary et al. 1993; Vogels and Orban 1996; Mysore et al. 2006). Yet, as a result of our strategic choice of stimuli, based on previous experience, the shape rank curves were rather steep (Fig. 4), providing a sensitive assay of selectivity. Another restriction was that we used only 5 rather simple geometric shapes. Interestingly, cells showing stronger shape selectivity tended to have more strongly conserved shape preferences across conditions. Thus, owing to the small set of shapes tested, it is possible that the more weakly selective neurons were simply not very strongly selective for any of these particular 5 shapes, and hence, we may have underestimated the invariance of the V4 neurons. Finally, we tested only one speed, 3 degrees/s, chosen so that acuity would still be high and the responses of low-pass neurons (Orban et al. 1986) would be strong. How invariance holds up as speed of shape motion increases remains to be explored.
Thus, our results clearly indicate that the ventral stream has evolved mechanisms for analyzing objects’ shapes that fit the requirements of the dynamic world. We have demonstrated that shape processing can use motion to break camouflage and extends to slowly moving objects in general. This invariance is robustly present at a relatively early level of the ventral pathway. That V4 neurons are selective for features more complex than lines but simpler than global shapes ensures that the invariance will be inherited by the downstream areas such as TEO and TE, at which level full shape selectivity emerges.
Different Cues in Kinetic Boundaries
The motion-defined shapes utilized in the current study (KS and KM conditions) contained a temporal frequency cue that is lacking in the omKS stimuli used previously in single-cell studies (Sary et al. 1993, 1995; Marcar et al. 1995, 2000; Mysore et al. 2006). This temporal frequency cue has been shown psychophysically to be a strong visual cue for boundaries (Sary et al. 1994; Nawrot et al. 1996) and may explain why shape selectivity was more pronounced for the KS than for the omKS condition (Table 3). Direct comparison of the 2 kinetic conditions reveals stronger average responses and shorter latencies for the KS than for the omKS conditions, as might be expected for conditions containing a more salient cue. What is more unexpected regarding the differences in salience is that the presence of this temporal frequency cue in the kinetic boundaries also promotes a higher level of invariance in shape selectivity. Among the V4 population that was shape selective in all 5 shape conditions, the interaction between shape rank and shape condition was stronger between LS and omKS than between LS and KS. This observation was also confirmed by the analysis of the separability values between pairs of conditions.
The temporal frequency cue is identical to the temporal texture cue used by Albright (1992) in his MT/V5 study. Selectivity for motion direction in MT/V5 neurons is largely invariant across cues defining the moving stimulus, whether luminance or temporal texture (Albright 1992; Olavarria et al. 1992). This finding represents the symmetrical result for the motion pathway that is analogous to our observation of invariance for shape selectivity in V4. It suggests in turn that invariance with regard to the defining cue might be rooted in properties of the afferents to MT/V5 and V4. In fact, studies in early visual areas (Chaudhuri and Albright 1997; Bourne et al. 2002; Lui et al. 2005) have shown that neurons in these areas already exhibit substantial invariance for both orientation and direction selectivity when temporal texture and luminance are used to define moving bars. In these studies, the responses to the 2 cues were rather different, in that the responses to bars defined by temporal texture were weaker than those to luminance-defined bars and the onset of responses to motion-defined bars was later than the onset of responses to luminance-defined bars (Bourne et al. 2002; Lui et al. 2005). The authors attributed this difference to the higher salience of the luminance cue. In our V4 study, the response levels to kinetic and luminance-defined shapes were rather similar, for both static and moving boundaries. This probably reflects the presence of additional cues in our kinetic stimuli and also the fact that we reduced the luminance contrast by adding a small proportion of randomly placed pixels of the opposite polarity in our black and white shapes.
Chaudhuri and Albright (1997), following the suggestion of Chubb and Sperling (1988), suggest that temporal texture stimuli could be detected by a mechanism that relies upon rectification of temporally modulated inputs, in other words some sort of flicker detector. Some evidence indicates that this hypothesized nonlinearity forces a partial equivalence of luminance and temporal contrast in the magnocellular laminae of the Lateral Geniculate Nucleus (LGN) (Derrington and Lennie 1984). Indeed, it has been demonstrated that projections of magnocellular afferents not only reach MT/V5 (Maunsell et al. 1990) but also V4 (Ferrera et al. 1994).
In this scheme, the invariance for luminance-defined and kinetic shapes would be of completely different origin for the KS, which includes an additional temporal frequency cue, and the omKS shapes. The former invariance might well originate in the LGN. This source is then very similar to that of the other aspect of the invariance we documented here in V4, that for LS and LM shapes. The origin of this invariance is traceable back at least to the LGN and has been observed from the earliest studies in V1 (Hubel and Wiesel 1961, 1967; Schiller et al. 1976). On the other hand, the invariance for the omKS might well originate in V4 itself (Mysore et al. 2006), using the few direction-selective neurons present in V4 as input. This hypothesis is supported by the difference in the response latencies of V4 neurons for LS and omKS (Mysore et al. 2006), which has been reconfirmed in the present study. It is not surprising that responses to omKS stimuli emerge at a later stage because they require the extraction of motion direction that does not occur until V1. Thus, the ethologically more important invariance would arise earlier in the visual processing stream.
Our results clearly demonstrate that at the level of V4, shape processing is largely independent of the cue defining the shape, luminance, or motion against a static background and that this cue-independent selectivity is preserved even when the boundary defining the object is moving. In contrast to invariance for cues requiring extraction of attributes such as motion direction or disparity, this ecologically crucial invariance may well be rooted in the properties of neurons at much earlier level in the visual system.
Supplementary material and supporting text can be found at: http://www.cercor.oxfordjournals.org/.
Flemish Regional Ministry of Education (GOA 2005/18); Queen Elizabeth Medical Foundation (GSKE); Katholieke Universiteit of Leuvon EF/05/014.
We thank M. De Paep, P. Kayenbergh, G. Meulemans, and Y. Celis for technical support. We also thank Dr M. Missal (UCL) for assistance with the eye velocity analysis. Conflict of Interest: None declared.