Multidimensional signal detection decision models of the uncertainty task: Application to face perception

Abstract

The uncertainty paradigm has been used in vision research to evaluate whether stimulus components are processed independently or not. The paradigm consists of several experimental conditions from which sensitivity indices are estimated and combined to provide evidence for or against the independence of stimulus components in perception. In typical applications, a multicomponent stimulus differs in one of its components from a standard value and the observer needs to decide if the change is an increment or decrement. In the certainty condition, the observer knows which component will contain the change; in the uncertainty condition, the component that differs from standard is unknown. Performance across the two conditions can be compared to that which is predicted by independence of components. The mathematical foundations upon which performance indices are related to component independence have been inadequately examined in previous applications and we clarify many of these concepts here. We derive predictions for observer sensitivity in the uncertainty condition and a relative measure, root-mean-square (RMS) that incorporates both uncertainty and certainty performance for three major decision models using a signal detection theory framework: a distance-classifier, the optimal decision model, and a decisionally separable (“independent” decisions) strategy. We also consider, using these decision models, implications for sensitivity and RMS when stimulus components are perceptually correlated. We present data from an experiment involving the perception of facial features in order to demonstrate how to apply the theoretical results.