Abstract
Human performance in matching unfamiliar faces to images is known to be poor, with far-reaching implications in a variety of workplace settings. Luckily, expertise for recognizing and matching particular individuals is highly trainable — through exposure, we can learn to distinguish even identical twins. Popular training methods ask participants to repeatedly perform a matching task or recognition task using photographs of the target identities, with feedback given after each trial. However, by limiting the training set to only original photographs of the target individuals, many of these training tasks are too easy and result in inefficient perceptual learning, with performance quickly hitting the ceiling. Here, we compare the effectiveness of more challenging face-training regimens that focus on the differences among a set of target faces to be learned. To accomplish this, we make use of recent advances in machine learning that provide the capability to encode photographs into a learned face space and then generate photorealistic morphs that interpolate between mid-level features of the depicted individuals, such as the taper of the chin and bushiness of the eyebrows. In particular, we used the StyleGAN neural-network architecture to generate several challenging variants of the Glasgow Face Matching Test (GFMT). In each variant, we morph each face towards the centroid of the set of all GFMT faces in the learned face space. We then empirically compare the efficacy of these GFMT variants for training face matching, measuring transfer of learning from the trained set to the original faces.