Multiple object tracking explained with neither fixed nor flexible resources
33.4028, Sunday, 17-May, 8:30 am - 12:30 pm, Pavilion
Zheng Ma1, Sheng-hua Zhong1, 2, Colin Wilson3, Jonathan Flombaum1; 1Department of Psychological and Brain Sciences , Johns Hopkins University, 2Research Institute of Future Media Computing, Shenzhen Univerisity, China, 3Department of Cognitive Science, Johns Hopkins University
In the multiple object tracking task an observer tracks moving targets within a set of featurally identical objects. Because error rates increase as target load (number of tracked targets) increases, the task is frequently cited as evidence of commodity-like resources that limit attention and working memory, with ongoing debates focused on whether resource allocation is fixed or flexible. Here we challenge the common assumption that resources limit tracking performance in the first place, adducing evidence from eye tracking experiments and a probabilistic model. The model infers the positions of targets on the basis of noisy samples received at a rate of 12Hz (similar results obtain at 20Hz), and it uses a nearest-neighbor algorithm to address correspondences between tracked targets and packets of unlabeled samples (including from nontargets). The noise in received samples depends on the distance of the relevant object from fixation (following Bouma’s law). As a result, fixation position greatly impacts performance. 20 participants completed tracking trials with loads between three and eight and several speed settings. The model then simulated each participant, assuming his or her actual fixations. Without fitting parameters and with the same settings (apart from fixations) under all conditions, the model produced speed and load effects that were significantly correlated with those of human observers. The model also captured trial-by-trial performance variance (controlling for load). And most strikingly, it accounted for between-subject variance, ranking participants as they ranked in practice. In summary: A model with a limited sampling rate and eccentricity dependent noise —but with neither fixed nor flexible resource limits— produces the typical load and speed effects taken as evidence of resources. Incorporating inter-observer fixation differences accounts for inter-observer differences in performance. These results suggest that MOT capacity limits are effective, not inherent, arising because of interactions between task parameters, fixation, and correspondence computations.