Abstract
The left and right sides of the brain differ in how they process visual stimuli such as words & faces, local & global aspects of Navon stimuli, high & low frequency gratings, even illusory contours. Our "differential encoding" autoencoder model suggests these differences can be explained by an asymmetry in the average length of long-range lateral connections, which are known to be involved in contour and shape processing. Our model accounts for more types of stimuli than competing models (e.g. the Double Filtering by Frequency (DFF) model), and makes a testable prediction about neural differences underlying lateralized visual processing. Here we account for two important phenomena our model has not been tested on. Two experiments show lateralization in coordinate and categorical discriminations. Coordinate discriminations require judging continuous distances and show right hemisphere (RH) dominance; categorical discriminations simply require classification of visual stimuli and show left hemisphere dominance (LH). Another two experiments examine response changes after "contrast balancing" stimuli. There, dark lines are outlined by thin bright lines, suppressing low frequency information while preserving the percept of shape and contour. Contrast balancing abolishes the "global precedence" effect (where the global level in Navon stimuli is processed faster than the local level), and suppresses most (not all) asymmetries. Across these experiments, our model shows LH dominance for categorical tasks and RH dominance for coordinate tasks. This asymmetry in our model is abolished by contrast balancing. Beyond these general patterns, our model closely matches the patterns for specific stimuli in each experiment. In addition, contrast balancing abolishes asymmetric processing of Navon stimuli in our model and removes the global precedence effect. Compared to the DFF model, our model fits the data better, accounts for more data, and uses a mechanism that could explain how asymmetry can survive contrast balancing. We plan to examine such data next.
Meeting abstract presented at VSS 2016