Here we present results regarding the spatial pattern of the fMRI selectivity for 3 initially novel object classes (smoothies, spikies, and cubies, see ) across IT cortex. First, we used data collected from the very first day the subjects saw the object classes and standard methods to reveal that IT cortex contains a subset of voxels that “individually” show reproducible, highly significant selectivity among the 3 novel object classes. However, we found that these voxels are just the tip of the iceberg, and we proceed to show that, using correlation methods on the same data, IT cortex contains distributed and reproducible spatial patterns of selectivity. Second, we performed a series of experimental manipulations to test the stability of this selectivity pattern across time and training, across behavioral task, and across changes in object position. Third, we tested manipulations of stimulus shape by changing the global and local shape properties of the object classes. Finally, we describe the relationship between this selectivity pattern and the previously described face patches in IT cortex. All experiments were designed, a priori, to target IT cortex, while using the earlier visual areas as controls. Nevertheless, as described in Supplementary Material
, we found no stable patterns of selectivity for these novel object classes in other brain regions that have previously been shown to prefer intact object images over scrambled images.
Reproducible Topography of Selectivity for Novel Object Classes across IT Cortex
In each of 2 monkeys that were naive to the object classes before the first fMRI scans, fMRI scanning revealed a reliable spatial distribution of object-class selectivity spanning IT cortex (Experiment 1). That is, different IT voxels preferred different object classes, a finding that we refer to as a “topography of selectivity” throughout this text.
We began our analyses by using standard methods to determine if any IT voxels were individually statistically selective among the 3 novel object classes. shows the significance maps across IT in each pairwise comparison of novel object classes. These maps illustrate that some voxels in IT are selective when considered individually (an average of 19% of IT voxels per pairwise comparison when thresholded at P < 0.001, uncorrected). We obtained an unbiased measure of the replicability and strength of this object-class selectivity across scan sessions (separated by 8 days on average) by performing test–retest analyses for voxels with strong object-class selectivity (see Materials and Methods). In particular, we selected IT voxels with a significant preference for one object class over another in 1 scan session (novel-object–selective ROI described in the Materials and Methods). We found that these same voxels showed a similar preference in other, non-overlapping scan sessions (), and this between-session replication was significant across independent time series in each monkey (J: P = 0.02; M: P < 0.001, see Materials and Methods). The response of each voxel to its nonpreferred object class was, on average, only 62% of its response to its preferred object class. Throughout this paper, we quantify this selectivity by the object-class selectivity index—the difference in the response to the preferred and nonpreferred object classes times 100 (after normalizing the response to the preferred object class to 1). In , the mean object-class selectivity index was 38% (the expected value if no selectivity existed is 0%).
In sum, we found reproducible selectivity in IT among initially novel object classes, at least when we focus on the voxels with the strongest selectivity (above). However, these regions might only be the “tip of the iceberg.” IT cortex might contain a reproducible pattern of selectivity that spans a larger region of IT than that revealed when voxels are only considered individually (Haxby et al. 2001
). To investigate this, we used correlation methods to analyze the pattern of selectivity across all voxels in IT cortex with a preference for objects over scrambled images (the standard-object–selective ROI; average of 1264 voxels per monkey; see Materials and Methods). Specifically, in each monkey, we first computed 3 pairwise response difference maps for these voxels (i.e., 1 map for each comparison: [smoothies – spikies], [smoothies – cubies], and [cubies – spikies]) and then computed the reliability of each pairwise map by computing the correlation of these maps across nonoverlapping halves of the data from Experiment 1 (see Materials and Methods). The reliability was 0.57 (standard deviation [SD]: 0.12) in monkey J and 0.73 (SD: 0.08) in monkey M (SD computed across different divisions of the data in halves). This significant reliability (P
< 0.01 in each monkey) shows that IT cortex contains a highly reproducible pattern of selectivity among the novel object classes. This reliable pattern of selectivity did not depend solely on the aforementioned voxels with significant selectivity when considered individually (the voxels used for the test–retest analyses in ). In particular, when we removed these voxels from the correlation analysis, then the reliability of the selectivity map in the nonselective voxels was nearly as high (0.55, SD: 0.13 in monkey J, and 0.69, SD: 0.11 in monkey M; P
< 0.01 in each monkey). Thus, even across all these “nonselective” voxels, we found a statistically reliable pattern of selectivity at the voxel “population” level.
Because our correlation analysis revealed that the fMRI data contain more selectivity than was uncovered in , we used a color map method applied to unthresholded maps to illustrate the topography of selectivity across IT cortex. Each potential pattern of class selectivity is captured with a unique color, and illustrates how the color map was constructed from pairwise t
-maps (also see Materials and Methods). Given that these maps are not thresholded, the goal of these color maps is not to show the significance of individual voxels but to visually illustrate the pattern of object-class preferences across IT cortex and its reproducibility across experiments (see further); in this sense, they resemble the color maps generated with optical imaging techniques. Thresholded maps that only reveal the “tip of the iceberg” of voxelwise reproducible selectivity are shown for each experiment in Supplementary Figure 3
shows the color maps for 1 hemisphere in each monkey together with the response to each object class for a few locations in the color map. These responses illustrate that the color maps represent relative preferences for specific object classes rather than a complete segregation of the voxels responding to different object classes. This observation is consistent with the test–retest analyses (see above) that showed that even the most selective voxels did not respond only to a single object class—voxels that preferred one object class also responded to other object classes (relative to a no-object fixation condition), albeit less strongly.
The Topography of Selectivity Is Stable across Time and Training
We scanned both monkeys again after a 3-month interval, using the same methods and behavioral task. During the 3-month interval, each monkey received extensive behavioral training in discriminating among exemplars within just 1 of the 3 object classes (monkey J trained with only spikies, monkey M trained with only smoothies; see Materials and Methods). The other 2 object classes were not seen during this interval. The goal of this training was to do our best to impart special meaning to just 1 of the object classes and to improve the monkey’s expertise within that class. We did not aim to distinguish among these and other potential effects of this behavioral training but first aimed to see if such training might alter each monkey’s IT pattern of object-class selectivity described above.
Stability across Time
We found that our fMRI measurements of object-class selectivity were remarkably stable across the 3-month training interval. The same voxels selected above for the test–retest analyses (selected using data before the training interval) had a mean selectivity index for the same object class that was nearly identical after the 3-month interval (34%) as computed before the interval (38%, see above; ). This selectivity index was highly significant in each monkey (J: P < 0.001; M: P < 0.001, t-test across time series), and in neither of the monkeys was it significantly different from the originally obtained selectivity index (J: P > 0.5; M: P > 0.15, t-test across time series). These numbers are averaged across all 3 object classes, trained and untrained, but similar results were found if the analyses (both for the selection of voxels with significant selectivity and for the computation of selectivity) were restricted to pairwise stimulus comparisons that involved the trained object class either as a preferred or a nonpreferred class (object-class selectivity index of 35% before training and 33% after training).
Although this analysis of selectivity indicates that the “macro” magnitude of selectivity for the novel object classes in the most selective voxels is roughly the same following the 3-month training interval, it does not provide direct insight into the stability of the spatial pattern of selectivity across IT cortex. To examine this, we plotted the selectivity topography that we measured following the 3-month interval (using the same color scheme as in ), and we found that the color maps were strikingly similar to those obtained 3 months earlier (). To quantitatively compare the spatial pattern of selectivity across the training interval, we computed the spatial correlation of the selectivity topography maps. Specifically, in each monkey, we again computed 3 pairwise response difference maps for all voxels in the standard-object–selective ROI and then computed the correlation of each pairwise map across the 3-month interval. The correlation (averaged across the 3 pairwise maps) was 0.51 in monkey J and 0.53 in monkey M (see ), and it was strongly significant in each monkey (J: P < 0.001; M: P < 0.001, t-test across time series, see Materials and Methods). The correlation across the training interval was also significant for each object pair map in each monkey (monkey J—[smoothies –spikies]: r = 0.72, P < 0.0001; [smoothies – cubies]: r = 0.44, P < 0.0001; [cubies – spikies]: r = 0.38, P < 0.0001; monkey M—[smoothies – spikies]: r = 0.56, P < 0.001; [smoothies –cubies]: r = 0.57, P < 0.01; [cubies – spikies]: r = 0.45, P < 0.001). Thus, we found a reliable pattern of selectivity for each of the 3 possible pairwise contrast maps. Furthermore, the average correlation (0.51) was not significantly smaller than the maximum correlation of 0.55 that is expected if the spatial distribution of selectivity was not changed at all during the interval of 3 months (J: P > 0.4; M: P > 0.4; the expected correlation takes into account the variability in each data set, see Materials and Methods and bar plots in ). In sum, within the limits of our data, we found that the spatial distribution of activity across IT cortex was remarkably stable across the 3-month training interval.
A drawback of both the analysis of the macro magnitude of selectivity (selectivity index) and the spatial correlation analysis described above is that multiple scan sessions (days) were needed to provide reliable results (the analyses rely on across-session replication of selectivity). Thus, such analyses do not necessarily rule out changes in selectivity that might have happened very rapidly (e.g., over a period of a few days), and for most of our results, the denotation of the 3 object classes as “novel” should thus be taken as a relative concept with this time scale in mind. Furthermore, only the performance of monkey M was good enough in the first scan session with the novel objects to include the data in the analyses (monkey J did not fixate well enough in this first session; see Supplementary Material
). Thus, it is possible that these analyses missed relatively rapid changes in the topography of selectivity for novel objects. However, more specific analyses suggest that is unlikely. In particular, the selectivity topography computed from only the first scan session with these novel objects in each monkey (i.e., in monkey J using data with relatively poor fixation performance—around 80%) was already highly correlated to that obtained after the 3-month interval (r
= 0.58 on average; cf., r
= 0.52 for all data from Experiment 1, see above). This suggests that the selectivity topography shown in and was already present on the first day that the monkeys saw the novel objects.
Stability across Training
To further investigate possible effects of shape discrimination training, we also focused on the spatial distribution of activity for the trained object class, independently from the spatial distribution of activity for nontrained objects. We used the responses to the object-class cubies as a baseline because this class was not trained in either monkey. This provided us with 2 measures of selectivity per voxel, 1 for trained objects (the response to the trained class minus the response to the cubies class) and 1 for nontrained objects (the response to the other class minus the response to the cubies class; for monkey J: smoothies [nontrained] – cubies [nontrained]; for monkey M: spikies [nontrained] – cubies [nontrained]). With this procedure, the trained and nontrained object classes were perfectly counterbalanced across monkeys. There was no consistent difference in the replicability of the spatial distribution of selectivity for the trained class compared with the nontrained class—monkey J: trained r = 0.38 and non-trained r = 0.44; monkey M: trained r = 0.57 and nontrained r = 0.45. Thus, the spatial distribution of activity associated with initially novel objects was as stable across months for an object category that was not seen during this time interval as it was for an object class that was shown more than 130 000 times in a task in which the within-class shape differences were learned.
This far, all analyses focused on the pattern of responses across IT cortex but have ignored potential differences in the “grand mean” of the response to trained and nontrained objects. We tested for training-related changes in overall response in the standard-object–selective ROI with a 2-tailed unpaired t
-test across time series that compared the posttrained difference in response between trained and nontrained objects with the pretrained difference in response between trained and nontrained objects (to control for possible preexisting differences in responsiveness not due to training). This analysis revealed no significant effect of training on the response to the trained object class in IT cortex (standard-object–selective ROI)—J: P
= 0.28; M: P
= 0.27; t
-test taking the data of the 2 monkeys together: P
= 0.10 (). Nevertheless, there was a small trend toward stronger responses for trained than for nontrained objects that was also seen in other brain regions (primary visual cortex and prefrontal cortex; see Supplementary Fig. 4
). These analyses focus on relative differences between trained and nontrained objects (normalizing to the nontrained responses), but the absolute, unnormalized responses were very similar before and after training (see caption of ). Furthermore, measurements of sensitivity for one object class versus another, normalized to the SD rather than to the mean response, also revealed no differential effect of training on the sensitivity for trained compared with nontrained objects (see Supplementary Fig. 6
). In summary, training did not measurably affect the spatial distribution of selectivity, and it had no significant effect on the overall response to the trained object class.
The Novel Object Selectivity Topography Was Tolerant to Changes in Behavioral Task
We found that the spatial distribution of novel-object–evoked activity was not only stable across time and training but also stable across different task contexts. During the collection of the fMRI data described so far, the shape of the objects was irrelevant to the monkeys’ task—the monkeys were rewarded only for fixating and for reporting color changes (color task; see Materials and Methods). After the 3-month interval, we also scanned each monkey during performance of the very demanding within-class shape discrimination task in which they were trained (shape discrimination task; see Materials and Methods). We found little effect of this task change on the selectivity topography. For example, an examination of the voxels with significant pairwise selectivity for at least 1 object class (the same novel-object–selective ROI as used above) revealed nearly the same magnitude of selectivity during the shape discrimination task (31%) as already described in the data obtained during the color task (see ). This selectivity was highly significant in each monkey (J: P < 0.01; M: P < 0.005) and in none of the monkeys was it significantly different from the selectivity observed during performance of the color task before the 3-month interval (J: P > 0.5; M: P > 0.5). The correlation in pairwise selectivity between the color task and the shape discrimination task across all voxels of the standard-object–selective ROI was 0.33 averaged across monkeys and object pairs (). This correlation was significantly positive in each monkey (J: P < 0.001; M: P < 0.05). Although it tended to be smaller than the maximum correlation of 0.42 that is expected given the variability in the data, this effect was not significant in either monkey (J: P > 0.4; M: P > 0.1). Thus, the topography of selectivity noted in a color task was largely replicated in a different task context.
The small difference in responsivity to trained versus nontrained objects that was not significant in IT cortex during the color task (described above) was more apparent during performance of the shape discrimination task ()—monkey J: P = 0.039; M: P = 0.099; taking all data of the 2 monkeys together: P = 0.010. The size of this training-related response increase was not significantly different between the color task and the shape discrimination task (P = 0.20 with all data of the 2 monkeys together). This response increase is visible in , where the color associated with the trained object class (green and red for monkey J and M, respectively) tends to dominate the color maps obtained during the shape discrimination task, much more than in the color maps obtained prior to training during the color task (compare top row and third row of ). In sum, the combination of training to discriminate among objects within 1 object class and performance of that shape discrimination task resulted in somewhat higher fMRI responses to the trained class relative to the other, nontrained classes, while the spatial distribution of selectivity was largely unaffected by these factors.
The Novel Object Selectivity Topography Was Tolerant to Changes in Object Position
Although the previous analyses show that different object classes produce reliable, stable patterns of IT selectivity in at least 2 tasks, they leave open many important questions about what aspects of the object classes produce those patterns. A possible candidate is the retinotopic envelope of the 3 object classes. Even though we jittered the stimulus position, each object category was still associated with a specific retinotopic envelope (see Materials and Methods and Supplementary Figs 1
). The potential importance of retinotopic envelope was confirmed by positive spatial correlations in pairwise selectivity in early retinotopic visual cortex (parafoveal V1 ROI) across the 3-month time interval (average correlation of 0.41, significant in each monkey, J: P
< 0.001, M: P
Because neurophysiologic studies have shown that IT neurons show some tolerance to changes in object position (Kobatake and Tanaka 1994
; Ito et al. 1995
; Op de Beeck and Vogels 2000
), we investigated the effect of changes in object position on the patterns of novel object-class selectivity uncovered here. Specifically, we attempted to isolate any position-tolerant component of the object-class selectivity by measuring the selectivity topography not only when objects were at the fovea with the original small scatter in object position (as already described) but also in separate scanning blocks in which we more strongly varied object retinal position.
In a first test, we presented the stimuli at eccentric visual field positions so that they were completely nonoverlapping with the visual field positions covered during the foveal scan sessions (specifically, objects were centered at 8.5-degree eccentricity in the 4 visual quadrants). In each monkey, we found that this change in stimulus position greatly reduced responses across ventral visual cortex compared with a control scan with foveal stimuli. In monkey J, the primary visual cortex response (parafoveal V1 ROI, see Materials and Methods) to these nonfoveal objects was 11% of the response to foveal objects, and the IT response was 38% of the response to foveal objects. In monkey M, the same position test had an even stronger effect on responsiveness (reduction to 9% and 26% of the response to foveal stimuli in V1 and in IT cortex, respectively). This strong effect of eccentricity on response strength supports findings in the literature that processing in ventral visual cortex is strongly biased toward foveal stimuli (Op de Beeck and Vogels 2000
; Brewer et al. 2002
). Given the substantially weaker responses to nonfoveal objects, it was not surprising that the pattern of selectivity in IT cortex (across all voxels in the standard-object–selective ROI) for 8.5-degree eccentric stimuli was only weakly reliable in monkey J (reliability of 0.29; see Materials and Methods), and not reliable in monkey M (reliability of 0.07). Because of this result, we tested the effect of changes in object position in monkey M by using jittered foveal stimuli in which the retinotopic envelope for each object class created by that jitter was altered and broadened compared with the small jitter used to obtain our original data (see Supplementary Fig. 1
). Simulations confirmed that this change in stimulus position would completely alter the pattern of response in a retinotopic map (see Supplementary Material
and Supplementary Figs 1 and 2
). Thus, the following analyses were applied to data obtained with 8.5-degree eccentric stimuli in monkey J and to more foveal but highly position-jittered stimuli in monkey M.
Examination of IT voxels with significant pairwise selectivity for at least 1 object class using the original object position (the same novel-object–selective ROI as used above) revealed nearly the same selectivity index in blocks in which object position was varied (object-class selectivity index of 36%, see ). This selectivity was significantly different from 0 in each monkey (J: P < 0.01; M: P < 0.05) and in neither monkey was it significantly different from the selectivity in the originally tested (foveal) position (J: P > 0.5; M: P > 0.35).
In IT cortex (standard-object–selective ROI), the across-voxel correlation in pairwise selectivity computed from objects presented in the original position and computed from objects presented at the new positions was 0.44 averaged across monkeys and object pairs (). This correlation was significantly positive in each monkey (J: P
< 0.001; M: P
< 0.001) and not significantly smaller than the maximum correlation of 0.49 expected given the variability in the data (J: P
> 0.4; M: P
> 0.30). For comparison, we also performed the same analysis on V1 voxels (parafoveal V1 ROI) and found that the correlation in pairwise selectivity between the original position and the new position was only 0.04 (P
> 0.3 in each monkey). The same result was found in early visual cortex as a whole—that is, all the visually active cortex ventral to the STS and more posterior than (not including) the estimated location of TEO (monkey J: 2450 voxels; monkey M: 6470 voxels). In this large region of “pre-IT” cortex, the initial pattern of selectivity as found with the original object position was replicated after a 3-month interval when using the same object position (monkey J: r
= 0.27, P
< 0.001; monkey M: r
= 0.46, P
< 0.0001) but not when stimulus position was changed (monkey J: r
= 0.03, P
> 0.30; monkey M: r
= 0.09, P
> 0.15). This result is illustrated in . Unlike the largely stable pattern seen in IT, the strong effect of the object position changes on the observed object-class selectivity pattern in early visual cortex is fully consistent with the differences in retinotopic envelope between object classes and experiments and their expected effect on the activation of retinotopic maps (as simulated in Supplementary Material
and Supplementary Fig. 2
Thus, the tested changes in object position largely preserved the spatial distribution of object-class selectivity in IT cortex, but not in retinotopic cortex. This indicates that the object-class selectivity that we observed in IT cortex was largely tolerant to changes in the retinal position of the objects, at least as long as the objects were not presented too eccentrically (in which case much lower responses were noted in each monkey and even a disappearance of the selectivity topography in 1 monkey).
The Novel Object Selectivity Topography Was Altered by Changes in Object Shape
Taken together, the data presented thus far are most consistent with the hypothesis that the patterns of object-class selectivity in IT reflect spatially varying preferences for what is broadly referred to as object “shape.” Although these data are not by themselves sufficient to determine exactly what aspects of shape are represented, the fact that shape selectivity is found for novel objects opens the possibility of investigating the determinants of this selectivity in a more systematic way than has been done before with familiar object classes. To both confirm that shape is an important factor (unlike time, task, training, and retinal position) and to take a first step in this direction, we tested simple variations in object shape. In particular, we created new object classes () that were derived by altering the original object classes in the combination of global shape features (including aspect ratio) and local shape features (e.g., whether a object contains many straight lines with right corners, many spikes ending into acute angles, or many smoothly changing curves). We scanned the same 2 monkeys using these 3 new object classes to determine the impact of these shape manipulations on the patterns of selectivity described above. We merely use the labels “global” and “local” as summary labels for a series of shape changes, and we do not claim that these are the most important dimensions of shape. Instead, our main goal was to show that the selectivity topography and our methods are sensitive to changes in shape.
We found that manipulation of global and local shape features produced systematic alterations in the patterns of IT shape selectivity. As in the analyses described above, we computed the across-voxel correlation in pairwise selectivity, here taking pairs of objects that corresponded in either global or local shape. For example, we correlated the pattern of responses in the [smoothies – spikies] contrast with the [spiky smoothies – cuby spikies] contrast (corresponding global shape) and with the [smoothy cubies – spiky smoothies] contrast (corresponding local shape). Correlations tended to be positive for both global shape (r = 0.29) and local shape (r = 0.13) comparisons (). Although these correlations tended to be significant in the individual monkeys (global properties: J: P < 0.001, M: P < 0.03; local properties: J: P < 0.10; M: P < 0.03), they were often significantly smaller than the maximum correlation of 0.55 that can be expected given the variability in the data (global shape: J: P > 0.4, M: P < 0.001; local shape: J: P < 0.001, M: P < 0.001). Thus, in contrast to our previous manipulations of time, training, and object position, these shape manipulations had a clear impact on the spatial patterns of IT selectivity. Moreover, neither of these 2 shape properties was able to fully explain the original selectivity pattern, suggesting that the underlying IT functional organization reflects both global and local shape features.
The Relationship between the Selectivity Topography Uncovered with Novel Objects and Face Patches
Because previous fMRI studies in both humans and monkeys have revealed class-selective responses for familiar object categories (Spiridon and Kanwisher 2002
; Tsao et al. 2003
; Pinsk et al. 2005
), we wondered how the selectivity topography uncovered with novel objects used here might be related to those earlier results. For example, if previous studies using face stimuli were uncovering selectivity that reflects the shape properties of faces, in addition to familiarity or meaning, then one might expect some relationship between the spatial patterns of object selectivity reported here and the spatial distribution of face patches, for example, a correlation across voxels between the preference for faces over other objects and the preference for one of the novel object classes.
To examine this relationship, we first replicated previously reported fMRI results of face-selective patches in an around monkey IT cortex (Tsao et al. 2003
; Pinsk et al. 2005
) by comparing 3 stimulus conditions: rhesus monkey faces, well-known objects (mainly fruits, toys, and lab apparel), and Fourier-scrambled images of these objects. shows color maps of the functional organization the same region of IT cortex for these 3 stimulus conditions. The face “patches” in these images are seen as red/orange regions in these maps, and their location is consistent with previous work (Tsao et al. 2003
). The pattern of face selectivity across IT cortex was very reliable across time series, with a reliability (see Materials and Methods) of 0.70 and 0.85 in monkey J and M, respectively. This is higher than the reliability of the pairwise comparisons between novel object classes (0.57 and 0.74 in monkey J and M).
A qualitative comparison of the spatial location of the face patches in each monkey with the novel-object selectivity topography () reveals a relationship between the 2 selectivity patterns. In particular, the face patches tended to be found in or around regions that prefer smoothies (and to a lesser extent cubies) above spikies (pink regions on the original maps). To quantify this relationship, we computed a pairwise selectivity index for faces, [faces – objects], and we correlated this index with the 3 pairwise selectivity indices that were introduced before: [smoothies – spikies], [smoothies – cubies], and [cubies –spikies]. The absolute value of most of these spatial correlations was significantly above 0 (P < 0.05; the only exception was the correlation with [cubies – spikies] in monkey M with P < 0.10), with an average correlation of 0.31 across object pairs and across monkeys (see ). This analysis confirms that there is at least a partial relationship between the spatial distribution of novel objects selectivity and the distribution of face selectivity. However, each of these correlations was also significantly smaller (P < 0.01) than the maximum correlation (0.71 on average) that can be expected given the high reliability of the data. Thus, a substantial proportion of the distribution of novel object selectivity cannot be simply reduced to the distribution of face selectivity (and vice versa).
This conclusion was confirmed when we predicted the selectivity for novel objects after a 3-month interval from the originally measured novel object selectivity and from face selectivity. We performed a multiple regression analysis with the selectivity in each voxel in a pairwise comparison after the interval as the dependent variable and 3 predictors: the selectivity in the same object pair before the interval, face selectivity, and a constant to estimate the offset. This analysis was done for each pairwise comparison in each monkey (n = 6), and each time we found a larger standardized regression coefficient for the novel object selectivity (average coefficient 0.49; different from 0 across the 6 analyses: P = 0.0002) than for the face selectivity (average coefficient 0.12; not different from 0 across the 6 analyses: P = 0.12). This difference in predictive value between novel object selectivity and face selectivity was significant across voxels in each analysis (P < 0.001), and it was significant across the 6 analyses (t(5) = 3.99, P = 0.010). Thus, most of the temporally stable selectivity topography for novel objects could not be reduced to the distribution of face selectivity.
Finally, we found very replicable patterns of selectivity for novel objects across regions of IT cortex that showed no preference for faces over objects. The aforementioned correlation of the novel object selectivity over the 3-month time interval (average correlation of 0.52) was obtained using the data of all object-selective voxels in anterior IT cortex (standard-object–selective ROI). If the reproducible novel object selectivity reflects more than face selectivity, then a sizable correlation should be obtained if the voxels contained in the face patches are eliminated from the previous analyses. Indeed, even after removing all IT voxels with any face selectivity (even if nonsignificant, 19% of standard-object–selective ROI removed), the selectivity topography in the remaining voxels was still highly correlated across the 3-month interval (0.47 on average; significantly higher than 0 for each object pair, P < 0.05).
In sum, the novel object classes revealed reproducible functional organization in IT outside of the face patches, and similarly, the distribution of face patches could not fully explain the pairwise selectivity across any of the novel object classes. However, the partial relationship between the selectivity topography for novel object classes and the spatial distribution of face patches suggests that at least part of the latter is due to shape properties of faces that are also found in initially novel objects.