Article Figures & Data
Figures
- Figure 1.
Example images of the six natural scene categories used in this study (from top to bottom): beaches, buildings, forests, highways, industry, and mountains, as well as four randomly selected examples of the perceptual masks used in the behavioral experiments (bottom row).
- Figure 2.
Confusion matrices for behavioral responses (left) and decoder predictions of fMRI activity in PPA (right). The rows of this matrix indicate the scene categories presented to the subjects (ground truth), and the columns the subjects' behavioral response (left) and the predictions by the decoder (right). An ideal confusion matrix would have 1 everywhere on the diagonal (correct classifications) and 0 in the off-diagonal entries (errors). Frequent confusions are highlighted in yellow.
- Figure 3.
Correlations of error patterns. The 30 off-diagonal entries (errors) in the fMRI-decoding confusion matrices (Fig. 2) are plotted over the errors in the behavioral experiment. The dashed lines show least-squares fits of linear relationships. Agreement of the error patterns was assessed with the Pearson product–moment correlation coefficient. High correlation was found for PPA, RSC, and LOC. *p < 0.05, **p < 0.01, †p = 0.069.
- Figure 4.
Effects of scene inversion. A decoder that was trained on fMRI activity from upright scenes showed significantly lower accuracy when decoding fMRI activity from inverted (light gray bars) than upright (dark gray bars) scenes in V1,PPA, and, marginally, RSC. When comparing the upright decoder tested on activity from upright scenes (dark gray bars) with a decoder that was trained and tested on fMRI activity from inverted scenes (white bars), only PPA showed a significant decrease in decoding accuracy. Error bars are SEM over five subjects. *p < 0.05, **p < 0.01, ***p < 0.001, †p = 0.052, and ††p = 0.072.
- Figure 5.
Axial (a) and sagittal (b) view of the whole-brain searchlight analysis. Areas in yellow show decoding accuracy significantly above chance (p < 0.01, corrected at the cluster level). Localizer-based ROIs from a single subject are marked in red, and overlap between the searchlight result and ROIs is shown in orange. In addition to visual areas around the PPA and the right fusiform gyrus (rFs), the inferior frontal gyrus (IFG) showed decoding accuracy significantly above chance.
Tables
ROI Decoding accuracy Error correlation Image similarity correlation Inversion effect V1 26%* 0.21 0.46** 0% FFA 22% 0.10 0.03 2% LOC 24%* 0.42* −0.22 3% RSC 27%* 0.34† −0.24 2% PPA 31%** 0.57** −0.07 7%* -
Decoding accuracy is measured in percentage of blocks predicted correctly, and significance is assessed relative to chance (17%). Error correlation establishes a correlation between misclassifications (off-diagonal entries in the confusion matrices) (Figs. 2, 3) between decoding from ROIs and human behavior. Image similarity correlation correlates the image similarities matrix with the confusion matrix from fMRI decoding. The inversion effect is defined as the difference in accuracy of a decoder trained and tested with upright versus trained and tested with inverted scene presentations. PPA shows significant effects in all analyses except for the image similarity correlation.
-
↵*p < 0.05;
-
↵**p < 0.01;
-
↵†p = 0.069.
-
ROI Mean SD V1 12.3% 3.7% FFA 6.5% 14.6% LOC 3.4% 3.3% RSC 60.0% 13.8% PPA 39.7% 13.1%
