Efficient task generalization and humanlike face perception in models that learn to discriminate face geometry

Artificial deep neural networks (DNNs) can excel at face recognition from 2D photographs where both shape and appearance cues abound; however, DNNs have rarely been challenged to recognize faces strictly based on face geometry. Here, we show that DNNs, even those fine-tuned on face photographs, had almost no generalization performance to a new geometry-based face task, while in the opposite direction, networks fine-tuned only on geometrically defined, textureless faces readily generalized to textured faces. To learn geometry in a more practical setting with colored and textured faces, we trained discrimination on face emotion in addition to face identity, which resulted in less texture bias and generalized well across face tasks. Learning in this way from just four individuals and their expressions generalized to unseen individuals, even exceeding standard models which are trained on classifying hundreds of face identities. Compared to standard models, emotion and identity trained models developed more humanlike errors in the identities or emotions that they confused. This novel method learns in a humanlike manner using only a few individuals but enriched with expressions that widely vary face geometry - similar to early human experience during child-parent interactions. Thus, this bioinspired work has broad implications for how moving toward humanlike learning of geometry in artificial vision can be both highly sample efficient and highly performing.