Learning Generalisation to Morph Faces: Connections to Memory Theory

Memory models, such as dual-process and global matching models, have been developed to explain patterns recognition memory. However, there have been conflicting results from studies using different stimuli and little exploration of the models’ ties to findings in social cognition. The current study investigated the relationship between model assumptions, the transfer of associated information from learned faces to perceptually similar faces (learning generalisation), and false recognition of two faces blended into a mathematical average (50/50 morphs).

In two experiments, participants learned faces associated with categories (friend or enemy). Participants then gave honesty and recognition ratings for old faces, morph faces, and new faces. In Experiment 1, participants rated congruent morphs, which blended two learned faces (parents) with compatible categories (e.g., friend and friend). In experiment 2, participants rated incongruent morphs, which blended two parents with opposing categories (e.g., friend and enemy), and partial morphs, which blended a learned parent (friend or enemy) with a new face.

Across the experiments, the results produced evidence of a small learning generalisation effect where the morphs’ honesty ratings were affected by the parents’ associated category. The honesty ratings of incongruent morphs indicated that the opposing categories balanced out. However, not all statistical analyses were consistent with expected learning generalisation patterns. Multiple conflicting effects, such as the level of associated learning or a familiarity bias, may have influenced honesty ratings and limited the effect of the associated categories.

The similarity of the morphs to the parents affected recognition ratings. All morph types were falsely recognised at a higher rate than new faces. Across experiments, congruent and incongruent morphs were falsely recognised at the same rate as old faces were correctly recognised and at a higher rate than partial morphs, which was predicted. The high rate of false recognition for 50/50 morphs was likely due to morphs retaining both featural and configural (perceptual) information from the parents. There was no consistent relationship between learning generalisation and recognition response, although the lack of a significant interaction may reflect an item selection issue and requires further investigation.

Different memory models can account for patterns in learning generalisation and false recognition of morph faces, but each model struggles to account for more complex outcomes. Modifications may overcome those issues, with the Retrieving Effectively from Memory (REM) model showing the most promise. Overall, the current study bridged research areas and demonstrated the efficacy of a memory-based approach to learning generalisation.