The Effects of Target Typicality on Guidance and Verification in Categorical Search
33.42, Sunday, May 12, 8:30 am - 12:30 pm, Orchid Ballroom
Justin T. Maxfield1, Westri Stadler1, Gregory J. Zelinsky1,2; 1Department of Psychology, Stony Brook University, 2Department of Computer Science, Stony Brook University
How are effects of target typicality expressed in a categorical visual search task? Participants were cued with a basic-level category name, followed by a present/absent search display depicting five real-world objects. Images of targets were segregated into interleaved high-, medium-, and low-typicality conditions based on typicality ratings (1-7) obtained during pretest; distractors were from random non-target categories. Contrary to previous work (Castelhano, Pollatsek, & Cave, 2008), we found a significant effect of target typicality on guidance; targets were first fixated most frequently in the high-typicality condition, followed by medium- and low-typicality targets (all p <.05). As expected, typicality effects were also expressed in target verification; the time between first fixation on the target and the manual response was shorter for higher typicality targets (all p <.05). To test whether these typicality effects on guidance and verification might be mediated by the same underlying process, we trained linear SVM classifiers on the same target categories used in the search experiment. Using SIFT features and positive and negative samples from ImageNet, we obtained for each of the high/medium/low typicality targets probability estimates of category membership based on distances from the SVM classification boundaries. We found that mean classifier confidence not only increased with target typicality, but these mean classifier estimates also correlated perfectly with the average (over participants) time-to-target (r = -.999, p = 0.028) and target verification (r = -1, p = 0.004) measures obtained for each level of typicality. Our findings clarify previous work by showing effects of target typicality on both search guidance and verification. Moreover, the fact that a single classifier could be trained to predict these components of search further suggests that the target representation used to mediate guidance and recognition search decisions may be one and the same.