Object Representations in Human Visual Cortex Formed Through Temporal Integration of Dynamic Partial Shape Views

Stimuli.

The stimuli were abstract, parametrically defined 3D shapes with complex shape characteristics (Fig. 1A) generated by a parametric shape model (“superformula”; Gielis, 2003). Two model parameters were varied to create a 2D parameter space. Nine combinations of the parameters (i.e., nine points in the space arranged in a cross-like pattern) were chosen to create nine different shapes. The shape images were then equated (without background) in terms of mean luminance and contrast.

Experimental paradigm.

We used an event-related fMRI paradigm. In slit runs, the slit orientation was either vertical or horizontal (Fig. 1B–D). The shapes were not shown before as a whole. In each trial, participants viewed one of the shapes moving behind a narrow slit in a fixed speed (6.6°/s) orthogonally to the slit orientation. The shape initially appeared in the slit after a 0.083 s delay, moved across the slit for 0.783 s until its edges disappeared, and then (after additional 0.15 s delay) reversed its direction and completed the move across the slit in another 0.783 s. The initial direction of the shape motion was chosen randomly. The active phase of the trial lasted 1.8 s followed by 1.2 s of “empty slit.” Participants were required to fixate a fixation point. To maintain attention to the stimuli and to ensure shape perception, they were also asked to press a button whenever the same shape appeared in two consecutive trials (a one-back recognition task). The same shape was shown in 8.4% of such trials.

The slit width was set to 0.47° of visual angle. The shapes subtended 4.7 × 4.7° of visual angle. Therefore, only 1/10 of the shape contour was visible at any moment through the slit. The fixation point, a small red rectangle, was placed to the left/right of the vertical slit and on the top/bottom of the horizontal slit so that it would not interfere with the motion within the slit. Half of the participants fixated on the “left” and “top” fixation point in the vertical and horizontal runs and the rest fixated on the “right” and “bottom” fixation point. Each of the nine abstract shapes was presented three times over the course of either vertical or horizontal run in a counterbalanced manner using the Optseq procedure. The trials (3 s) were pseudorandomly embedded with fixation periods (an “empty slit” with a fixation point) ranging from 1 to 9 s. Each participant completed eight vertical-slit and eight horizontal-slit runs. The runs were acquired in an interleaved order (vertical, horizontal, vertical, horizontal, etc.). Each run was 115–119 s long (115–119 volumes) and ended with a 10–12 s fixation period. The total amount of trials in slit runs comprised 3 repetitions × 9 shapes × 16 runs = 432 trials.

Only after all slit runs were completed were subjects presented for the first time with full images of the shapes and performed the same task. In each trial of a full-image run (Fig. 1E), a full shape image was shown for 0.8 s followed by 1.2 s fixation period (a screen with a central small red rectangle). Each of the nine shapes was presented four times in each run. The total number of trials in full-image runs comprised 4 repetitions × 9 shapes × 8 runs = 288 trials. Eight full-image runs were completed in a counterbalanced manner. As above, the trials (2 s long) were pseudorandomly embedded with fixation periods ranging from 1 to 10 s. Each run lasted 113–120 s (113–120 volumes) ending with a 10–12 s fixation period. The total number of trials in full-image runs comprised 4 repetitions × 9 shapes × 8 runs = 288 trials.

Eye tracking.

Participants were required to maintain their gaze on a fixation point throughout the experiment. To ensure that fixation was maintained in the slit fMRI trials, the monocular eye position of five (of 16) participants was tracked during three to six slit runs using a video-based, infrared illumination eye tracker (Eye Link1000; SR Research) with a sampling rate of 250 Hz. To determine fixation gaze stability, we computed spatial dispersion of eye position (Di Russo et al., 2003) across the active phase of a slit trial (when the shapes were moving). Because visual motion can skew gaze position (Zimmermann et al., 2012), the dispersion may be larger in the axis of motion direction (i.e., in X for vertical-slit trials and Y for horizontal-slit trials). The dispersion was small relative to the slit width (0.47°) ranging across participants between 0.16° and 0.43° (mean 0.26°) for the motion axis dimension and between 0.11° and 0.35° (mean 0.26°) for orthogonal dimension. This indicates that fixation was relatively stable in both cases. A 2-way ANOVA on the dispersion values with two factors slit orientation and dispersion dimension confirmed that spatial dispersion of eye position did not differ between the dimensions (F_(1,4) = 0.21, p = 0.667 and F_(1,4) = 0.002, p = 0.970, respectively).

Stimuli.

The same shapes served as stimuli.

Experimental paradigm.

Before and after the fMRI scans, participants performed a similarity-rating task under slit-viewing conditions. In the first rating session (before fMRI scans), the participants had not seen the whole shapes yet. The experiment started with a very short familiarization phase in which each of the nine shapes were shown once (translating behind a vertical slit). All participants reported that they clearly perceived a whole shape moving behind a slit. This ensured that they were able to perform the subsequent experiments (i.e., reliable similarity rating in the behavioral experiment and a one-back shape recognition task in fMRI scans). Next, they were asked to rate shape similarity using a rating scale of 1–7, where 1 was the most dissimilar and 7 was the most similar. Subjects were instructed to use the entire scale for their responses and to rate shape similarity taking into account the 3D shape structure. In each trial, the two shapes were seen sequentially through a slit and then participants reported their degree of similarity by pressing a keyboard button. An example of such a trial is shown schematically in Figure 2A. The shapes moved exactly as in the fMRI trials: for 1.8 s each with interval of 1.4 s between them. Each trial ended after a response from the response button. Vertical- and horizontal-slit trials were presented in two different blocks to prevent generalization of shape information across slits. Each of the 45 shape pairs (36 pairs comprising images of different shapes and nine pairs with images of the same shape) was presented six times in a pseudorandom order across the blocks (three times per block). All other trial details were the same as in the fMRI main experiment.

In the second rating session, the participants had already seen the whole shapes in the course of the fMRI experiment. This time, to reinforce manifestation of global shape percept, the first shape in a pair was shown through a vertical slit and the second shape was presented behind a horizontal slit (and vice versa in other trial; Fig. 2B). All other experimental details were the same as in the first rating session. Each session lasted ∼35 min.

Statistical analysis.

First, we determined the reliability of the scores; that is, the extent to which our rating procedure yielded the same result for each shape pair on repeated trials in the same participant and across participants. We calculated both individual and intersubject reliability with a split-half method and a Spearman–Brown correction (as commonly accepted in psychometrics; Carmines and Zeller, 1979). Specifically, we split the total number of pair assessment trials to two halves, averaged the scores per each half, and correlated the resulting two vectors of scores (using Pearson correlation). We then applied the Spearman–Brown formula [2 * r/(1 + r)] to yield the reliability estimate for the full set of trials (we performed the split 100 times to get a good estimate of reliability). The obtained reliability coefficients were further averaged across participants. To estimate the intersubject reliability, we split the total group of subjects in two halves and then proceeded as above. The within- and across-subject reliability was 0.94 and 0.98, respectively, indicating the high overall consistency of the similarity rating scores (reliability estimates vary within the range between 0 and 1).

In further analysis, the similarity-rating scores were averaged across trials per shape pair for each session in each participant. We normalized them to values between 0 and 1 and converted these values to perceived shape dissimilarity estimates (by subtracting the resulting values from 1). Each session's estimates for 36 pairs (comprising images of different shapes) were used, resulting in two dissimilarity vectors. Next, we assessed the degree of perceptual distance correspondence between the sessions (before and after seeing the full images) by calculating the Pearson correlation between the two dissimilarity vectors in each participant. The resulting r values (one per subject) were Fisher z-transformed to allow averaging across subjects. We then calculated the mean and SE and determined the lower and upper bounds of the 95% confidence interval. The mean and the bounds were transformed back into correlations (which are reported). Furthermore, to assess the degree of perceptual distance correspondence within the group, we calculated the correlation between the dissimilarity vectors (pooled across the two behavioral sessions) across participants, resulting in the group mean and its confidence intervals. We also computed a pixelwise Euclidean distance measure of shape similarity (Grill-Spector et al., 1999). Perceptual distances between the shapes assessed by the similarity-rating task were tested for their correspondence to the pixelwise physical distances in each individual.

Stimuli.

The rationale behind this experiment was to ensure that shape recognition was not based on a distinct feature allowing for shape recognition without a need for a temporal integration process. To test this explicitly, we presented the same set of slit-dependent shape views in a random temporal order, thereby eliminating the possibility of temporal shape integration but preserving all features within a given frame. The same shapes served as stimuli at the same rate and during the same trial periods.

Experimental paradigm.

The experiment included vertical and horizontal slit conditions only. However, the shape views (slivers) were now presented in a shuffled order, thereby eliminating the percept of both global shape and coherent motion (Morgan et al., 1982). The exact shuffle order was varied from trial to trial. To maintain attention to the stimuli and their shape features, participants performed a one-back recognition task: they were asked to press a button whenever the same set of shape views appeared in two consecutive trials. Each participant completed eight vertical-slit and eight horizontal-slit runs with the same parameters as above. The task was perceptually difficult, but it was performed well above chance [group mean d′ (SD) = 1.0 (0.6)].

LOC localizer.

All participants completed the LOC localizer scan. Grayscale images of real objects/abstract shapes and phase-scrambled versions of these images (while preserving the original power spectra of the images) were presented in a block design fashion. The localizer scan was composed of four conditions, objects, abstract shapes, phase-scrambled objects, and phase-scrambled abstract shapes, with 11 blocks for each condition. Each block lasted 8 s and was composed of 10 images (0.7 s per image with a 0.1 s interval of gray screen with a fixation point). The images were selected of a pool of 45 photographs for each category (real objects and abstract shapes) and were presented centrally, spanning 9 × 9°. Participants were instructed to fixate on a central fixation point and to press a button whenever the same image appeared twice consecutively. This occurred on average once per block.

Early visual cortex mapping.

We defined the early visual areas in all participants using two polar angle-mapping scans (Sereno et al., 1995; Engel et al., 1997). The polar angle stimuli consisted of a clockwise or counterclockwise rotating wedge composed of monochromatic checkerboard patterns with counter-phasing flicker frequency of 7.5 Hz. The radial size of the pattern segments was adjusted to approximate the cortical magnification factor (Sereno et al., 1995). Each wedge covered 45° of arc and extended from the fixation point to 8.5° into the visual periphery. Each scan, either clockwise or counterclockwise, comprised six cycles of the rotating wedge of 24 s duration. Participants were instructed to fixate on a central fixation point. To ensure that fixation was maintained throughout the run, the color of the fixation point changed for 0.1 s 9 times per scan, and participants were asked to indicate the change with a button press.

Visual motion localizer.

All participants completed the MT+ localizer scan (Huk et al., 2002). Participants were presented with a dot pattern that was either moving at 8°/s alternating direction radially inward and outward from fixation once per second (10 s) or stationary (10 s). The task was to fixate on a central fixation point and covertly count the number of fixation point blinks (presented 25 times throughout the scan). During a motion block, 1200 white dots on a black background were presented within a 16°×16° aperture centered at the fixation point. Each dot appeared for 0.1 s and was then replaced by another dot at a randomly selected position. The moving/stationary pair of blocks was repeated 12 times.

Body/face localizer.

A subset of our participants (n = 11) also completed the body/face localizer scan. This included three experimental conditions (headless bodies, faces, and everyday objects) with nine blocks for each condition. Each block lasted 10 s and was composed of 12 images (0.7 s per image with a 0.13 s interval). The three conditions were counterbalanced and interleaved with “baseline” blocks, including a fixation point only. The images were selected of a pool of 58 photographs for each category and were presented centrally, spanning 9 × 9°. Participants performed the same one-back task as above.

Standard ROI set.

Eleven bilateral ROIs were defined for each of the 16 participants. These were volume-based ROIs with 3 mm isotropic voxel resolution restricted to the gray matter. Figure 3 shows the ROIs of one representative participant. Six visual retinotopic areas (V1, V2, V3, V3ab, V4, and LO-1) were identified based on retinotopic mapping. Using a phase-encoding approach (Sereno et al., 1995; Engel et al., 1997), we constructed individual polar maps and overlaid them on inflated cortical surfaces (as implemented by BrainVoyager QX software). The retinotopic borders were delineated manually (Wandell et al., 2007) in each hemisphere. The volume-based ROIs for V1, V2, V3, V3ab, hV4, and LO-1 in both hemispheres were then extracted comprising 394 (95) [group mean (SD)], 309 (90), 250 (46), 241 (65), 97 (29), and 93 (43) voxels, respectively. MT+ was defined based on its greater BOLD response to moving dots compared with stationary dots. A false discovery rate (FDR) criterion was applied to correct for multiple comparisons. MT+ comprised significant contiguous voxels [q(FDR) < 0.05] located in the posterior inferior temporal sulcus (Huk et al., 2002). The group mean ROI size for MT+ was 164 (47) voxels across both hemispheres.

Object-selective regions were identified using our LOC localizer (Malach et al., 1995). To better separate between LOC and object-selective foci in the dorsal visual pathway (as well as between different LOC subparts), we first created three bilateral anatomical masks matching the standard anatomical boundaries for these regions (Grill-Spector et al., 1998; Golarai et al., 2007; Vinberg and Grill-Spector, 2008; Julian et al., 2012). The masks included the following: (1) the lateral occipital cortex, (2) the posterior and midfusiform gyrus, and (3) the caudal part of the intraparietal sulcus (cIPS). They did not include retinotopic cortex and MT+. LOC was identified within the lateral occipital cortex and the fusiform gyrus in both hemispheres (i.e., within both first and second anatomical masks) using the standard contrast: intact images > their scrambled counterparts, with [q(FDR) < 0.05] (Grill-Spector et al., 1999). To avoid issues related to the variability of LOC sizes across participants, we limited the number of voxels in the LOC ROIs: we sorted the voxels according to their t-values (in the standard LOC contrast) and picked the 350 voxels with highest values (if an individual LOC comprised <350 voxels, all of these voxels were taken for the analysis). This number was chosen conservatively based on LOC sizes (in cubic millimeters) reported previously in the literature (Golarai et al., 2007; Vinberg and Grill-Spector, 2008; Julian et al., 2012). LOC voxels were selected based on their t-values regardless of their contiguity (as long as they were within the standard anatomical boundaries); in most cases, they were contiguous.

In addition, using the same contrast (intact images > scrambled images), we identified two LOC subregions: the lateral occipital (LO) and posterior fusiform (pFs) regions (Grill-Spector et al., 1998; Vinberg and Grill-Spector, 2008). For each participant, we first selected significant voxels [q(FDR) < 0.05] located within the corresponding anatomical mask (either first or second). Then, the number of voxels was limited to the 200 or 150 most responsive voxels in LO or pFs, respectively, taking into account their reported sizes (Vinberg and Grill-Spector, 2008). Finally, object-selective regions in the dorsal visual pathway were also identified (using the third mask) and included cIPS and sometimes the adjacent transverse occipital sulcus (Grill-Spector et al., 1998, 2000).

Because LOC overlaps with category-specific visual regions (Downing et al., 2007), we defined a set of additional bilateral ROIs in 11 of the 16 participants using our body/face localizer. These included extrastriate body area (EBA) and fusiform body area (FBA) (Downing et al., 2001), as well as fusiform face area (FFA) and occipital face area (OFA) (Kanwisher et al., 1997). EBA and FBA were identified as lateral and ventral occipital regions, respectively, that were activated more strongly to images of human bodies than to objects [q(FDR) < 0.05]. FFA and OFA were defined in their standard anatomical locations based on their greater BOLD response to faces compared with objects [q(FDR) < 0.05; OFA was identified in 10 participants]. Voxels overlapping with the previously delineated retinotopic cortex and MT+ were excluded. The group mean sizes for the initially defined EBA, FBA, FFA, and OFA were 192 (31), 48 (30), 79 (30), and 51 (25) voxels, respectively. Six new ROIs were then constructed. First, in each participant, we labeled all mutual voxels between the LOC ROI and each body/face ROI. These voxels (comprising on average 29% from the total LOC volume) were grouped into “shared LOC” ROIs. The rest of LOC voxels represented “nonshared LOC” ROIs. EBA, FBA, FFA, and OFA ROIs were eventually restricted to the voxels that do not overlap with LOC.

Fixed-size ROI set: equalizing the number of voxels in each ROI.

Because ROI size can affect the results of multivariate pattern analysis in the ROIs (Schreiber and Krekelberg, 2013; Walther et al., 2016), our fixed-size ROI set was created so that all areas would have an equal number of voxels. As above, we applied activation-based voxel ranking, but this time, we chose voxels that responded best to all slit-viewed shapes regardless of slit orientation. Before the procedure, V4 and LO-1 were combined (due to their small sizes in some participants). V2 and V3 voxels were also pooled into one ROI. Therefore, the new set comprised nine ROIs in each of our 16 participants: V1, V2 + V3, V3ab, V4 + LO-1, MT+, cIPS, LOC, LO, and pFs. The ranking procedure was applied iteratively. In each iteration, ROI data from one vertical-slit run and one horizontal-slit run were taken jointly. Then, a single t-value for each voxel was calculated in the contrast testing all shape conditions against all rest periods (regardless of a slit orientation). Next, we sorted the voxels according to their t-values and took the 100 most active voxels that responded best to all slit-viewed shapes (a few individual ROIs comprised <100 voxels; all of these voxels were taken for the analysis). The remaining data were used for multivoxel pattern analysis (MVPA). The procedure was repeated eight times (according to the number of slit runs): the first vertical and first horizontal run were taken for voxel selection in the first iteration, etc. Therefore, in each iteration, an independent dataset was used for voxel selection. Importantly, MVPA and feature selection were always performed on different datasets to exclude issues of circular inference. The MVPA results were averaged across the iterations.

Note that, in a subset of our participants (n = 11), six additional ROIs were defined: “shared LOC”, “nonshared LOC”, EBA, FBA, FFA, and OFA. The EBA, FBA, FFA, and OFA ROIs were pooled into one body/face ROI (due to their small sizes in some participants). Using the above voxels' ranking procedure, “shared LOC” and “nonshared LOC” ROIs, as well as combined body/face ROI, were similarly limited to 100 voxels that responded best to all slit-viewed shapes (regardless of slit orientation). Therefore, the new set comprised 12 ROIs in these participants.

Cross-validation scheme.

We applied a cross-validation approach to assess the degree to which the voxel patterns of activation convey shape information in the various slit conditions. To that end, we first partitioned our data. In the main ROI analysis (for the standard ROI set), all collected data samples were used for MVPA (comprising 3 viewing conditions × 8 runs = 24 runs in total). For classification analysis, each viewing condition dataset was split into two unequal, nonoverlapping subsets composed of six and two runs. The data from six runs were used to train a SVM model. The remaining data from the other two runs were used to test the SVM performance. The splitting process was performed multiple times to reduce the variance of the performance (using all possible permutations of a split of the eight runs into the two subsamples, n = 28).

In the correlation analysis, we divided the data using the classical split-half approach (Haxby et al., 2001; Chan et al., 2010) and the unequal-split approach as in our SVM analysis (see also Walther et al., 2016). Both split schemes gave similar correlation-based decoding results. We used here the unequal split to be consistent with the classification analysis so that the variance of both SVM and correlation results could be similarly attenuated by cross-validation (due to the same cross-validation scheme). The raw data from runs within each subgroup (of six and two runs) were then z-transformed and concatenated.

In our equal-size ROI analysis, we divided the row data using two iteration loops. In the “outer” loop, one vertical-slit and one horizontal-slit run was chosen for voxel selection (and therefore was excluded from MVPA). This was done iteratively using the eight possible choices of the one run (e.g., first vertical and first horizontal run were chosen in the first iteration, etc., as explained previously). The remaining seven runs were further divided to two subgroups for MVPA using the “inner” loop. In classification analysis, five runs were used for training the SVM model and the two other runs were used for cross-validation testing. In the correlation analysis, the raw data from runs within each subgroup (a block of five or two runs) were z-transformed and concatenated. The inner loop cross-validation procedure was repeated 21 times (thus including all possible partitions of seven runs into subgroups of five and two runs).

Shape discrimination analysis.

We had nine predictors (i.e., one per each shape image) to fit the data in each voxel. In the classification analysis, this was done separately for each run. In the correlation analysis, the predictors were computed per run block consisting of either six or two runs in the main ROI analysis (using the standard ROI set) or either five or two runs in the equal-size ROI analysis (using the fixed-size ROI set). We then applied a standard GLM to estimate the BOLD response per each shape in each separate run (in the classification analysis) or subgroup (in the correlation analysis). Before analyzing the resulting patterns of regression coefficients, we used a denoising procedure designed to account for the noise in the patterns of activation. This was done because noise normalization enhances the reliability of both pattern classification and the estimates of dissimilarity between the activation patterns (Walther et al., 2016). The noise structure in each voxel within a ROI was assessed from the residual error of the GLM. We then normalized each voxel's regression coefficient by the SD of its residual, thus suppressing the contribution of noisy voxels that can have high (either positive or negative) regression coefficients due to high noise (Misaki et al., 2010). In addition, we accounted for the noise covariance between the voxels in each ROI (for further details, see Walther et al., 2016).

To perform pattern classification, the regression coefficients were normalized by subtracting from each coefficient the population mean and dividing by the population SD. The SVM model used a linear kernel function and one-versus-rest multiclass classification scheme (Reddy et al., 2010). SVMs were trained iteratively (based on the above cross-validation scheme) to distinguish between the shapes in one viewing condition (e.g., vertical-slit) and its generalization performance was tested in the other two conditions (horizontal-slit or full-image runs and vice versa). In total, there were six combinations of viewing conditions. Classification performance across different slit-viewing conditions (two combinations) and across slit and full viewing conditions (four combinations) was averaged across the relevant combinations. Shape classification was also tested in full-image runs (to get a benchmark for performance) with generalization performance assessed across different runs.

Consequently, 3 9 × 9 confusion matrices were calculated by assessing the proportion of trials of each predicted shape (by the classifier) for each input shape: each row in the matrices represents a specific shape that was shown and each column represents a predicted shape. Therefore, each entry in the matrix P(i,j) is simply the proportion of cases in which the classifier classified shape i as being shape j. Correct classification accuracy is represented by the diagonal elements P(i,i) and was averaged to receive a mean value across all shapes. The accuracy was further averaged across the cross-validation iterations in each individual ROI. Chance level decoding is 1/9 = 11%. The original confusion matrices were used for further analysis (see next section).

In the correlation analysis, the response vectors per image were first extracted per iteration for each run subset and normalized by subtracting the voxel's mean activation level across shape-image conditions. In total, 18 response vectors (9 shapes × 2 subsets of the data) were extracted for each viewing condition (i.e., for vertical-slit, horizontal-slit, and full-image viewing). Next, for each pair of images, we calculated the Pearson correlation between their response patterns from different viewing conditions. That is, response vectors from the subset #1 (in one viewing condition) were cross-correlated with the response vectors from the subset #2 (in another viewing condition), resulting in 6 versions of 9 × 9 correlation matrix. Correlation values were Fisher's z-transformed. The two across-slit correlation matrices and four slit- versus full-viewing matrices were averaged. The benchmark correlation matrix was also calculated for the response vector pairs in the full-viewing condition across independent subsets of these vectors. Therefore, the end result was 3 9 × 9 correlation matrices: 2 of them measured the degree of similarity of the response vectors across different viewing conditions and 1 across different runs in the same full-viewing condition. To assess the degree of shape discrimination, we subtracted the mean off-diagonal scores (i.e., correlations between different-shape patterns) from the mean diagonal scores (i.e., correlations between same-shape patterns). The resulting excess correlation was averaged across the iterations in each ROI to obtain a shape discrimination score per participant. The original correlation matrices were used for further analysis (see next section).

Shape-related representational similarity analysis.

To assess the degree of correspondence between the similarity of the evoked brain responses to a pair of images and their perceptual resemblance (as reflected by their similarity rating score), we used representational similarity analysis (RSA). Perceptual distances between each two shapes (i.e., dissimilarity measures) were estimated in two separate behavioral sessions performed before and after scanning (as explained in “Behavioral experiment” section). The resulting estimates were highly correlated across the sessions (see Results). Therefore, the distances were pooled across the two sessions for each participant. Similarly, the distances were highly correlated across participants (see Results), so we decided to use their group mean average for the RSA.

To assess the degree of dissimilarity between brain activation patterns for each pair of shapes, the off-diagonal correlation scores in the correlation matrices (from the above discrimination analysis) were converted into dissimilarity measures. The dissimilarity (distance) was calculated as 1 − Pearson's r, yielding the representational dissimilarity matrix (RDM) (Kriegeskorte et al., 2008; Nili et al., 2014). Misclassifications (off-diagonal scores) in the classification matrices also provide information about the distance between patterns corresponding to different shapes (Haxby et al., 2014; Walther et al., 2016). We normalized the resulting values in the matrix (by dividing each entry by the maximum value in the matrix) and subtracted the resulting values from 1 (to get a pattern dissimilarity measure). Because our distance measures are derived after multivariate noise normalization and cross-validated, they are (relatively) independent of the patterns' noise and therefore reflect the true distances between the patterns (Walther et al., 2016).

Finally, we computed the Pearson correlation between the shape perceptual distances and response patterns' RDMs (Ejaz et al., 2015) per ROI. We then squared the resulting correlation coefficient to assess the proportion of variance in the RDMs (R²) that can be accounted for by the perceived shape dissimilarity.

ROI-based statistical analysis of shape information.

Statistical analysis was applied to the shape discrimination scores [i.e., classification accuracy (%) and excess correlation] and R² (i.e., the explained variance in the classification-based and correlation-based RDMs). In each ROI, for each generalization rule (e.g., generalization across different slit conditions), we determined whether the accuracy (across participants) was significantly greater than chance level using one-sample (two-tailed) t tests. The same approach was used to determine whether excess correlation and R² were significantly greater than zero. To differentiate between the various regions of interest (i.e., those showing view-invariant shape information vs those that do not have this capability) we applied a two-way repeated-measures ANOVA: The factors were ROI identity and the generalization capability. The generalization factor was a three-level factor that reflects the type of generalization: (1) across different slit-viewing conditions (2) across slit and full-viewing conditions, and (3) across different runs of the same full-viewing condition (Fig. 4, inset). The number of levels in the ROI identity factor was equal to the total number of ROIs (in either the standard or fixed-size ROI set), excluding LO and pFs (because the majority of LO and pFs voxels are contained within LOC ROI). Having found a significant interaction term using this analysis, we continued with a subsequent analysis of simple main effects using one-way ANOVA separately in each ROI. We also used post hoc paired (two-tailed) t tests between ROIs to test whether the shape information estimate (e.g., classification accuracy) was different between ROIs. Statistics were performed based on all participants in the study.

In the control fMRI experiment, we tested whether slit-invariant shape discrimination and R² change when the same set of slit-dependent shape views (slivers) are presented in a shuffled temporal order. A repeated-measures ANOVA with two factors, ROI identity (levels indicating different ROIs) and temporal order (two levels, indicating the correct or random temporal order) was run, followed up with post hoc t tests. Statistics were performed based on a subset of our participants (n = 11).

The simple main effects and t test statistics were corrected for multiple comparisons using Sidák's correction. In addition, to corroborate the robustness of the perceptual–neural correspondence estimate (R²), we ran a permutation analysis. The values within each previously calculated confusion matrix were permuted across the matrix space for each ROI in each participant. This procedure was repeated 2000 times and the same RSA analysis as described in the previous paragraph was repeated, resulting in a chance distribution of the group mean R² in each ROI. Given these chance distributions, we assessed the significance of the estimates. The obtained values were very close to the values from the parametric tests. We therefore reported the results of the parametric tests alone.

Whole-brain statistical analysis.

Statistical analysis was applied to cross-slit classification accuracy (percentage) per searchlight cluster and the explained variance in the cluster's classification-based RDM (R²). Clusterwise one-sample two-tailed t tests were used to identify search windows showing significant values of the estimates across 16 participants. The resulting t-value statistical maps were corrected for multiple comparisons [q(FDR) < 0.001]. For comparison, we present the location of the most significant clusters in the searchlight analysis and the location of the most significant ROI showing the greatest cross-slit shape sensitivity (LOC) (Fig. 10). To that end, we calculated the group probability map for LOC voxels. For this purpose, all individual LOC ROIs (from the standard ROI set) were combined into one ROI. The probability of each Talairach voxel was assessed by counting the proportion of individuals in which that specific voxel was included in the individual-based ROI.