Purchase this article with an account.
David Hoppe, Constantin Rothkopf; Modelling the dynamics of visual attention under uncertainty. Journal of Vision 2015;15(12):566. doi: https://doi.org/10.1167/15.12.566.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
While several recent studies have established the optimality of spatial targeting of gaze in humans, it is still unknown whether this optimality extends to the timing of gaze in time-varying environments. Moreover, it is unclear to what extent visual attention is guided by learning processes, which facilitate adaption to changes in the world. We present empirical evidence for significant changes in attentive visual behavior due to an observer's experience and learning. Crucially, we present a hierarchical Bayesian model, that not only fits our behavioral data but also explains dynamic changes of gaze patterns in terms of learning. We devised a controlled experiment to investigate how humans divide their attentional resources over time among multiple targets in order to achieve task-specific goals. Eye movement data was collected from the participants. During each trial, three stimuli were presented on a computer screen arranged in a triangle. Each stimulus consisted of a small dot moving randomly inside a circular boundary. Participants were asked to detect when a dot exceeded its boundary. Subjects showed significant differences in allocation strategies over time that led to behavioral changes. The hierarchical Bayesian model captures the course of these changes. For example, in accordance with our computational model participants switched between the targets at a higher rate if the underlying dynamics were unknown (early trials). However, as more information about the statistics underlying the movement of the stimuli was collected (later trials) the switching rates decreased. Our model provides an explanation for these changes by linking them to different stages of learning the dynamics of the experiment. The ideal learner proposed in this work extends the ideal observer underlying many computational approaches to understanding visual attention by taking hierarchical learning effects into account.
Meeting abstract presented at VSS 2015
This PDF is available to Subscribers Only