|
Abstract:
|
Nearly all imaginable human activities rest on a context -appropriate dynamic control of the flow of retinal data into the nervous system via eye movements . The brain’s task is to move the eyes so as to exert intelligent predictive control over the informational content of the retinal data stream . An intelligent oculomotor controller would first model future contingent upon each possible next action in the oculomotor repertoire , then rank -order the repertoire by assigning a value v (a ,t ) to each possible action a at each time t , and execute the oculomotor action with the highest predicted value each time . We present a striking evidence of such an intelligent neural control of human eyes in a laboratory task of visual search for a small target camouflaged by a natural -like stochastic texture , a task in which the value of fixating a given location naturally corresponds to the expected information gain about the unknown location of the target . Human searchers behave as if maintaining a map of beliefs (represented as probabilities ) about the target location , updating their beliefs with visual data obtained on each fixation optimally using the Bayes Rule . On average , human eye movement patterns appear remarkably consistent with an intelligent strategy of moving eyes to maximize the expected information gain , but inconsistent with the strategy of always foveating the currently most likely location of the target (a prevalent intuition in the existing theories ) . We derive principled , simple , accurate , and robust mathematical formulas to compute belief and information value maps across the search area on each fixation (or time step ) . The formulas are exact expressions in the limiting cases of small amount of information extracted , which occurs when the number of potential target locations is infinite , or when the time step is vanishingly small (used for online control of fixation duration ) . Under these circumstances , the computation of information value map reduces to a linear filtering of beliefs on each time step , and beliefs can be maintained simply as running weighted averages . A model algorithm employing these simple computations captures many statistical properties of human eye movements in our search task . |