Abstract
Ability to visually find targets is limited by noise inherent in the brain's sensory processing and random variability in the environment that can make an irrelevant distractor visually confusable with the sought target. In some occasions the variability is purely random and cannot be overcome even by an optimal detector. However, in other instances environments consist of target-like background structures that albeit contain some random components are not entirely unpredictable. Thus, in theory, unlike pure noise, these fortuitous target-like occurrences in the backgrounds could be potentially discounted if humans could learn to recognize and classify them as part of the environment. This learning could drastically increase the ability to find the targets. Here, we measure six observers' performance localizing Gaussian targets of random size and contrast embedded in white noise and a randomly spatially-shifted background composed of a fixed pattern of randomly configured target-like structures. Average human target localization performance was initially very poor (24 %) but then remarkably progressed (80 %) after 600 trials to exceed that of theoretical models that do not have knowledge of the background but are otherwise ideal (e.g, ideal observer in white noise with/without human visual contrast sensitivity). Humans achieved a maximum efficiency of 20 % relative to an ideal observer that uses full knowledge of the background to mitigate its detrimental effects on performance. Our results suggest that background learning can be an important mechanism to quickly adapt to the visual environment and thus maximize the probability of successful target search.
Support NIH EY015925, HL53455, NSF-0135118