Renninger, L.W., Verghese, P. & Coughlan, J. (2005). Eye movements can be understood within an information theoretic framework. Computational & Systems Neuroscience (Cosyne05).
Researchers have long sought to understand volitional eye movements. We move our eyes in order to collect high resolution samples with the fovea. How does an observer select a target for high-resolution scrutiny? This decision must be based on visual information in the periphery, knowledge of the stimulus, and the task or goals of the observer.
Information theory has served as a useful framework for engineering communication systems and can be applied as a mathematical tool to characterize biological systems. Information can be thought of as a measure of uncertainty in the system. In particular, the brain has been described as a hierarchical Bayesian inference machine (Lee & Mumford, 2002) whose main task is to reduce uncertainty about the environment. We focus on a core sensory process: vision, for which we a rich set of experimental tools exists. In this abstract, we use eye movement recordings, behavioral psychophysics and a computational model to ask to what degree eye movement planning can be understood as a process of sequential information maximization.
It is intuitive to think that we move our eyes to the most informative regions of an image. Previous experiments in reading have studied eye movements within an information theoretic framework. In particular, an ideal observer model clearly demonstrated how these concepts can be applied computationally (Legge, Hooven, Klitz,Mansfield & Tjan, 2002). Lee & Yu (2000) theorized that in fact, information available as the entropy of response distributions in V1 hypercolumns might be used for eye movement planning.
We have previously developed a model of human eye movements based on information maximization (Renninger, Coughlan & Verghese, 2005). Subjects were asked to study a novel shape silhouette and then discriminate it from a highly similar shape. The brevity of the study period and difficulty of the task forces subjects to be relatively efficient in their sampling behavior. Furthermore, the simplicity of the stimulus allowed us to model information as the entropy of a distribution over orientations at the edges. We demonstrate that subjects select highly informative regions for fixation, but that additional biological constraints must be included in the model to predict the sequence of these fixations.
We incorporate the falloff in resolution from the fovea to the periphery in our measurement of information, because this is a central motivation for moving the eyes. The eccentricity dependent measurement taken with each fixation can be modeled as a filtering of the image with orientation and scale selective filters (akin to simple cells in V1). Alternatively, we can utilize the notion of a perceptual hypercolumn to define an effective neighborhood over which orientation information is pooled (Levi, Klein & Aitesbaomo, 1985). We demonstrate that the two approaches produce similar representations of information.
We then ask whether or not this representation of information is perceptually relevant for an observer during the shape discrimination task. We compute a discriminability index (DI) as the difference in underlying information content for two similar shapes. One observer was asked to use a rating scale when selecting the shape studied from the highly similar pair, allowing us to disassociate confident answers from guesses. In trials where the subject is confident, the average DI is highest. The index is lower in trials where the observer is guessing correctly, and even lower when the observer is guessing incorrectly. Thus, our representation of information may be similar to that being used by observers in this task. We conclude that information theory provides a solid general framework for investigating the human oculomotor system.