Learning processes in perception
Lisbeth S. Fried · The Journal of the Acoustical Society of America · 1977
Signal detectability theory assumes observers who transform each observation into a form monotonic with likelihood ratio. Since evaluation of likelihood ratios presupposes knowledge of the stimulus distributions, preasymptotic variability in performance has been presumed to derive solely from shifts in decision criteria with feedback and not to shifts in the estimates of the distribution parameters themselves. These were assumed known [Dorfman and Biederman (1971); Dorfman, Saslow, and Simpson (1975); Kac (1962); Kubovy, Rapoport, and Tversky (1971); Kubovy and Healy (to be published); Larkin (1971); Schoeffler (1965); Thomas (1973, 1975)]. In this research a Gaussian Ideal Learner (GIL) is developed which estimates the five unknown parameters of two Gaussian distributions. Each observation is used to continuously update parameter estimates, while current parameter estimates are used to classify the observation. Estimates approach the true parameter values as n approaches infinity. Preasymptotic variability results from shifts in parameter estimates. Data from two experiments using normally distributed numerical stimuli and two using randomly distorted visual patterns are consistent with the model. Observers, operating without feedback, use the observations to estimate distribution parameters, and that preasymptotic variability is due to changes in parameter estimates and not merely to shifts in decision criteria. [Work supported by NSF Contract No. BNS-77-01211.]