Figure 5.

Network interactions between the early and late subsystems predict the decision bias toward face choices relative to house or car choices. A, Mean RT and precision for face and house choices, respectively. Precision is defined as the proportion of correct choices among all choices of a particular type. On average, face choices had faster RT (p = 0.0072, paired t test) and higher precision (p = 0.00074, paired t test) than house choices. Error bar represents the SEM across subjects. B, Performance measures of face choices for the face–house contrast. Subjects were divided into the High Face group (face precision > house precision) and the Low Face group (face precision < house precision). Compared with the Low Face group, the High Face group had lower sensitivity (p < 0.01) and higher specificity (p < 0.01), suggesting fewer false positives in their face choice, and therefore they were less biased toward faces. C, Performance measures of house choices for the face–house contrast. For subjects in the Low Face group, their house precision was relatively higher than the subjects in the High Face group not because they were better at detecting houses (indistinguishable sensitivity for houses); rather, it was because they were more biased toward faces and less likely to mistake a nonhouse for a house. The smaller number of false positives in houses for the Low Face group (higher specificity, p < 0.05) increased their house precision. D, Mean RT and precision for face and car choices, respectively. E, Performance measures of face choices for the face–car contrast, similar to B, subjects were divided into the High Face group (face precision > car precision) and the Low Face group (face precision < car precision). F, Performance measures of car choices for the face–car contrast, similar to C. G, Difference in early–late interaction is predictive of the difference in FPR (1 − specificity). Higher face FPR relatively to house is correlated with more network interactions (r = 0.84, p = 1.61 × 10⁻⁶). High face FPR indicated more bias toward faces. The bias toward faces was characterized by more interactions between the two subsystems. H, The same correlation between the network interactions and face bias holds for the face–car contrast.