Feature selection and utilization: a call for a two stage model of visual form perception

Abstract

A review of the psychophysics of form literature indicates that the multidimensional scaling (MOS) model accounts for perceptual judgments in a variety of perceptual tasks. The conceptual framework which incorporates the .MDS model appears, however, to be limited in scope. The limitation is the emphasis on the feature utilization processes to the exclusion of the feature selection processes. A general conceptual framework for a two stage model of perceptual judgments, incorporating both feature selection and utilization processes was proposed. The feasibility of such a two stage model was explored by examining the effects of three variables, task, stimulus format, and class structure, upon perceptual judgments. The stimuli used in the study consisted of ten eight-digit VARGUS 9 patterns (Evans & Mueller, 1966) drawn from each of two schematic classes and mapped into two stimulus formats, polygon and serriform. Eight Ss in each of the two stimulus format conditions performed in four perceptual tasks, repeated, in both the visual and memory modes, plus a paired-associate learning task. The four perceptual tasks included similarity estimation, magnitude estimation of the deviations of the forms from a specified form (the class prototype), recognition, and classification. In the visual mode, forms were used as stimuli, while in the memory mode, nonsense syllables were used as stimuli. In the memory mode, Ss were required to recall the forms previously associated with the nonsense syllable shown and to make judgments based on the remembered forms. Multidimensional scaling solutions (Kruskal, 1964) were found for the visual and memory similarity estimates and the visual and memory recognition latencies for both stimulus format conditions, polygon and serriform. Several important findings were reported. No differences were found in the number of recognition and classification errors across the polygon and serriform groups. Canonical correlations showed marked congruences in the scaling solutions across the four task and two stimulus format conditions. Psychophysical prediction of perceived differences in the spaces, using schematic and polygon physical measures was higher for the similarity spaces (R^2 > .70) than for the recognition latency spaces (R^2 < .50). INSCAL (Carroll & Chang, 1969) solutions for the combined serriform data and the combined polygon data resulted in three dimensional spaces with a between stimulus class and two within stimulus class dimensions for each data set. Comparisons of the INSCAL solutions for the two data sets showed correspondences in the relative importance of within and between class dimensions for the four types of task. And finally, prediction of the magnitude estimation data using dimensions in the scaling spaces resulted in multiple regression coefficients greater than .70 in value. In summary, these findings suggest that correspondences exist across tasks and stimulus formats in the stimulus attributes that humans use for perceptual judgments. Second, the results indicate that the distribution of stimuli on an attribute may serve as a basis for the selection of that attribute for use in a given task. And finally, the distributional characteristics of attributes in conjunction with task demands may determine, in part, the feature selection process in a two stage model of perceptual judgments.