Abstract
We investigated whether previous knowledge about the underlying structure of scenes influence eye movements during free exploration. Subjects (N=6) were presented with a sequence of scenes, each for 3 seconds, consisting of a 5x5 grid and 6 shapes in various cells of the grid. The scenes were composed of two triplet chunks (three element in fixed spatial relation) selected from an inventory of 4 triplets and configured randomly. Two hundred scenes were presented and eye movements of the subjects were recorded, while they performed a two-back comparison: they had to notice any change between the current display and the one before the previous display. Next, a 2AFC task was used to assess how well subjects learned the underlying statistical structure of the scenes by choosing the more familiar of the two presented pattern fragments in each trial. We trained an online probabilistic non-parametric ideal observer model to learn the underlying structure of the scene including the number and identity of chunks composing each scene. For each trial, we used the model to predict subjects' eye movements based on the estimated reduction in uncertainty about identities of shapes, given previous fixations in the scene and the knowledge of previous scenes. For each subject, we found a significant relationship between saccade length and the reduction of uncertainties produced by the saccade (p<0.001). While for very short saccades, there was no relation between model prediction and saccade direction (p>0.05), for longer saccades, the average uncertainties reduction was 0.355 (p<0.01) following the predictions of the model. Furthermore, there was a positive correlation between the success of prediction, and subjects' performance in the 2AFC test (r = 0.466). This provides evidence that eye movements during free exploration are influenced by implicit knowledge about the underlying structure of the scenes.