The long-term objective of the proposed work is to understand how learning by the visual system helps it to represent the immediate environment during perception. Because perception is accurate, we can know spatial layout: the shapes, orientations, sizes, and spatial locations of the objects and surfaces around us. But this accuracy requires that the visual system learn over time how best to interpret visual """"""""cues"""""""". These cues are the signals from the environment that the visual system extracts from the retinal images that are informative about spatial layout. Known cues include binocular disparity, texture gradients, occlusion relations, motion parallax, and familiar size, to name a few. How do these cues come to be interpreted correctly? A fundamental problem is that visual cues are ambiguous. Even if cues could be measured exactly (which they cannot, the visual system being a physical device) there would still be different possible 3D interpretations for a given set of cues. As a result, the visual system is forced to operate probabilistically: the way things """"""""look"""""""" to us reflects an implicit guess as to which interpretation of the cues is most likely to be correct. Each additional cue helps improve the guess. For example, the retinal image of a door could be interpreted as a vertical rectangle or as some other quadrilateral at a non-vertical orientation in space, and the shadow cues at the bottom of the door helps the system know that it's a vertical rectangle. What mechanisms do the visual system use to discern which cues are available for interpreting images correctly? The proposed work aims to answer this fundamental question about perceptual learning. It was recently shown that the visual system can detect and start using new cues for perception. This phenomenon can be studied in the laboratory using classical conditioning procedures that were previously developed to study learning in animals. In the proposed experiments, a model system is used to understand details about when this learning occurs and what is learned. The data will be compared to predictions based on older, analogous studies in the animal learning literature, and interpreted in the context of Bayesian statistical inference, especially machine learning theory. The proposed work benefits public health by characterizing the brain mechanisms that keep visual perception accurate. These mechanisms are at work in the many months during which a person with congenital cataracts learns to use vision after the cataracts are removed, and it is presumably these mechanisms that go awry when an individual with a family history of synesthesia or autism develops anomalous experience-dependent perceptual responses. Neurodegenerative diseases may disrupt visual learning, in which case visual learning tests could be used to detect disease;understanding the learning of new cues in human vision could lead to better computerized aids for the visually impaired;and knowing what causes a new cue to be learned could lead to new technologies for training people to perceive accurately in novel work environments.
Showing the most recent 10 out of 24 publications
