New perspective on face learning: Stability modulates resolution of facial representations in the optimal observer

Learning new faces is challenging, error prone, and subject to large individual differences. A dominant theory claims that we remember faces by creating averaged representations that discard irrelevant changes due to viewing circumstances and retain stable inner features (e.g., eyes). If so, why do we occasionally fail to recognise familiar people and why don’t we always encode inner features? We propose a parsimonious face encoding system, in which the relative stability of (extra)facial feature determines the resolution at which they are encoded, following a coarse-to-fine strategy. This is confirmed in an ecological learning paradigm where, all else being equal, faces with stable appearances were encoded more coarsely than faces with variable appearances. The framework assumes that individual differences are based on the ability to use cost-efficient encoding strategies flexibly. Accordingly, poor recognisers rigidly encode coarse features whereas a lack of stability encourages good recognisers to refine their representations.