Eye movements and the complexity of visual processing

We propose the hypothesis that the difficulty of a visual task for human observers can be estimated from its computational complexity, defined with respect to a specific constrained computer architecture. We suggest that eye movements may serve to reduce this constrained complexity for certain visual tasks. We begin by discussing human performance on several examples of perceptual tasks. We then informally introduce analysis of algorithmic complexity and examine the complexity of these tasks for parallel computational networks whose depth is constrained to a small number of levels. Finally, we discuss the ability of the human visual system to perform translation invariant pattern recognition. In accordance with predictions based on the complexity of visual tasks, eye movements can compensate for failures of the translation invariance of the visual system.