Uniting theory and data: the promise and challenge of creating an honest model of facial expression.

People routinely use facial expressions to communicate successfully and to regulate other's behaviour, yet modelling the form and meaning of these facial behaviours has proven surprisingly complex. One reason for this difficulty may lie in an over-reliance on the assumptions inherent in existing theories of facial expression - specifically that (1) there is a putative set of facial expressions that signal an internal state of emotion, (2) patterns of facial movement have been empirically linked to the prototypical emotions in this set, and (3) static, non-social, posed images from convenience samples are adequate to validate the first two assumptions. These assumptions have guided the creation of datasets, which are then used to train unrepresentative computational models of facial expression. In this article, we discuss existing theories of facial expression and review how they have shaped current facial expression recognition tools. We then discuss the resources that are available to help researchers build a more ecologically valid model of facial expressions.