Figure 4. Assessing the nature of the animacy continuum in the ventral temporal cortex (VTC).

(A) The region-of-interest, VTC, is highlighted. (B) The order of objects on the VTC animacy continuum, image categorizability (IC), perceptual categorizability (PC), and agency (Ag) are shown. (C) The within-participant correlations between VTC animacy and image categorizability (IC), perceptual categorizability (PC), visual categorizability (VC, a combination of image categorizability and perceptual categorizability; Materials and methods), and agency (Ag) are shown. All four models correlated positively with VTC animacy. (D) The left panel shows the correlations between VTC animacy and VC and Ag after regressing out the other measure from VTC animacy. Both correlations are positive, providing evidence for independent contributions of both agency and visual categorizability. The right panel shows the correlation between VTC animacy and a combination of agency and visual categorizability (Materials and methods). The combined model does not differ significantly from the VTC animacy noise ceiling (Materials and methods). This suggests that visual categorizability and agency are sufficient to explain the animacy organization in VTC. Error bars indicate 95% confidence intervals for the mean correlations.

Figure 4—figure supplement 1. The contribution of the principal components of VTC activations to VTC animacy.

Principal component analysis was performed on the average of the regression weights (VTC activations) for each of the 16 objects, which were computed using a general linear model for each run of the main experiment. (A) The average of the participant-wise percentage variance in the VTC activations explained by each of the 15 principal components is shown. (B) The average participant-wise correlations between the principal components (PCs) and VTC animacy are shown. Although the first PC captures more variance than the second PC, the second PC contributes more to VTC animacy than the first. None of the individual PCs can fully account for VTC animacy. (C) The average scores of the 16 objects on the first three PCs are shown. Animacy organizations are seen in all the three PCs, where humans and other mammals lie on one end, insects in the middle, and cars and other inanimate objects lie on the other end of the representational spaces.

Figure 4—figure supplement 2. The contributions of image and perceptual categorizabilities (IC and PC), independent of agency (Ag), to VTC animacy.

Agency is regressed out of IC and PC and correlations of those two residuals (IC_i and PC_i) with VTC animacy are shown in (A) and (B), respectively. Both IC and PC explain variance in VTC animacy that is not accounted for by Ag (p < 0.05 for both correlations). IC and PC are also separately regressed out of Ag and the correlations of the two residuals with VTC animacy are shown in (A) and (B) respectively (p < 0.05 for both correlations). Echoing the observations in Figure 4C, agency contributes to VTC animacy independently of either IC, PC, or a combination of both (visual categorizability - VC in Figure 4C).

Figure 4—figure supplement 3. The robustness of our findings to the choice of the layer of VGG-16 used to quantify image categorizability.

Eberhardt et al. (2016) showed that the image categorizability scores of objects (different images than the ones we used) from the second layer of the fifth group of convolutional layers (C5-2) of VGG-16 correlated highest with how quickly humans categorized those objects as animate or inanimate in a behavioral experiment. Results were highly similar when the image categorizability (IC) as used in Figure 4 was replaced by the image categorizability of C5-2. Specifically, the independent contributions of visual categorizability and agency to VTC animacy remained significant and the correlation between the combined model and VTC animacy was at VTC animacy noise ceiling.