Modular Serial-Parallel Network for Hierarchical Facial Representations

Researchers in the field of face perception have long debated the extent to which we process faces holistically (as a whole) versus analytically (focusing on individual features). While some evidence suggests that faces are perceived holistically, research on the neural organization of the visual system and people's subjective experiences indicate that face parts can also be analyzed individually. This view is corroborated by observations of hierarchical object representation in which selective neural populations are fine-tuned to detect specific visual properties ranging from simple features to more complex combination of features. Thus, advances in theories of face perception are met with two major challenges: How can hierarchical face representations be used to integrate holistic and analytic encoding within the same framework? And how can the stages of face processing be integrated with higher-level cognitive processes, such as memory and decision-making, that are recruited during facial perception? We propose a novel computational framework termed the Modular Serial-Parallel Network (MSPN), which synthesizes several perceptual and cognitive approaches including memory representations, signal detection theory, rule-based decision-making, mental architectures (serial and parallel processing), random walks, and process interactivity. MSPN provides a computational modeling account of four stages in face perception: (a) representational (b) decisional, (c) logical-rule implementation, and (d) modular stochastic accrual of information and can account for both choice probabilities and response-time predictions. In a face classification task, MSPN showed an impressive ability in fitting choice response time distributions over other models. MSPN can be used as a tool for further development and refinement of hypotheses in face perception. The analysis of the model’s parameter values, estimated from data, can be used to explore distinct properties of the perceptual and cognitive processes engaged in both analytic and holistic encoding. We conclude by outlining how MSPN could be generalized to other perceptual and cognitive domains.