Abstracts of Scientific Presentations 2019 AALAS National Meeting, Poster P211, Denver, Colorado
J Am Assoc Lab Anim Sci. 2019 Sep; 58(5): 607–726.
Face recognition is dependent on both the shape and appearance of faces, the former providing information about geometry and depth, the latter providing color and texture information. Previous face recognition research has often relied on photographs of frontal or profile views of faces, which demand little robustness across viewing conditions while confounding shape and appearance cues. In our study, to minimize the usage of texture cues and hone in on the effect of shape for face discrimination, we created 3D, textureless face meshes and rendered them in greyscale. In addition to measuring human performance on this task, we tested common marmosets, a New World monkey that has garnered attention as a small primate model for neuroscience. We sought to determine whether marmosets have the ability to recognize faces which has not been thoroughly tested but is ultimately critical to the marmoset’s utility as a model for studying the neurobiological mechanisms of human face processing.We trained three marmosets to discriminate between two different artificial face identities presented in varying pose, size, position, illumination, and background. On held-out images, marmoset performance was ~70% across different conditions (M1: 80%, M2: 75%, M3: 65%), compared to ~80% for humans. Above chance performance was nontrivial given the challenging nature of the task - various state-of-the-art deep neural networks (DNNs) which excelled at a basic-level object recognition control task (80-90%) performed quite poorly on our face discrimination task (50-60%). Rather than perform near chance, the same DNNs would typically perform near ceiling for face identification from photographs - stimuli that have also been used previously in face recognition researchBesides absolute performance, we are currently interested in whether marmosets demonstrate similar deficits in face processing as humans. One such deficit is the face inversion effect where humans have difficulty discriminating upside-down faces, more so than discriminating upside-down objects. In preliminary experiments, we found that performance for inverted faces was lower by 17% in one marmoset despite high proficiency at upright face discrimination (>80% performance), similar to the 14% performance decrement observed in humans. Our results show that marmosets can indeed perform fine geometry-based face discrimination that is challenging even for humans and machines. Furthermore, the presence of a face inversion effect, points toward the marmoset sharing a common repertoire for face recognition as humans, endorsing the argument for marmosets as a model for high-level visual neuroscience.
