Abstract
Why do we make eye fixations to specific places? The standard account of the locus of fixations has been that constellations of co-located features serve as the targets of gaze [1]. Although it has been suggested that fixation locations may be determined by a surprise measure that calculates whether the distribution of features at a location is unexpected [2], this account is still feature-based.
Our own hypothesis is that task-based reward governs the choice of locations [3]. Numerous studies have established that reward can be seen as the basis of eye fixations via correlations at the neural level, but we have described a completely novel hypothesis: Fixations can reduce uncertainty in the state of a cognitive behavior's control program. We have predicted that, given that the central evaluation is reward-based, eye fixations should be made to the location that promises the most reward for uncertainty reduction. This hypothesis had been tested in a virtual reality simulation with humanoid models and shown to be superior to standard methods.
We now report that our hypothesis has been tested using human subjects' walking data [4]. When subjects walk by approaching pedestrians, they fixate them with a probability that varies predictably depending on ancillary tasks that the subjects are engaged in, such as following a leader. Our reward based model predicts that these variations in probability will have a mean of 0.17 with a standard deviation of 0.03 which is very close to the observed mean of 0.19. The closeness of this fit suggests that the feature-based account of eye fixations may need extensive revision.
1. L. Itti and C. Koch, Nature Reviews Neuroscience, 2001
2. L. Itti and P. Baldi, IEEE CVPR, 2005
3. N. Sprague, D. Ballard and Al Robinson, ACM TAP, 2007
4. J. Jovancevic and M. Hayhoe, JOV, submitted