Abstract
Deep convolutional neural networks (DCNNs) have achieved human-level accuracy in face identification (Phillips et al., 2018), though it is unclear how accurately they discriminate highly similar faces. Here, we compared human and DCNN performance on a challenging face-identification task that includes identical twins. Participants (N=29) completed a face identification task in which they viewed pairs of high-quality, frontal images in three conditions: match (same identity, N=40), general imposter (different identities from similar demographic groups, N=40), and twin imposter (identical twin siblings, N=40). Ratings were collected on a 5-point Likert scale (1: sure different; 5: sure same). Performance was assessed for a DCNN trained for face identification (Ranjan et al., 2018) on the same image pairs used in the human experiment. Human and DCNN identification accuracy were measured using the area under the ROC curve (AUC). DCNN performance was similar to human performance when comparing the AUC for match pairs versus general-imposter pairs (human AUC=0.97, DCNN AUC=1.0). When comparing match pairs versus twin-imposter pairs, human performance declined substantially (AUC=0.79), but DCNN accuracy remained high (AUC=0.96). Human participants’ performance in the general-imposter condition correlated strongly with their performance in the twin-imposter condition (r=0.64). Next, we conducted an item analysis to examine the similarity of each image pair for both human participants and the DCNN. Rank-order correlation of these trials showed moderate correlation between human and DCNN rankings for match (⍴=0.50, p<0.001) and twin imposter (⍴=0.48, p<0.01) pairs, and no correlation for general imposter pairs. In conclusion, the DCNN outperformed human participants when distinguishing face images of identical twins. However, the relative difficulty of identifying high-resemblance faces was related across DCNNs and humans. This study provides insight into the accuracy of humans and DCNNs when discriminating high-resemblance face identities.