Abstract
Over the last 20 years, neuroimaging techniques have shed light on the modulatory nature of top-down feedback signals in the visual system. What is the functional role of top-down feedback and might there be multiple types of feedback that can be implemented through automatic and controlled processes? Studies of voluntary covert attention have demonstrated the flexible nature of attentional templates, which can be tuned to particular spatial locations, visual features or to the structure of more complex objects. Although top-down feedback is typically attributed to visual attention, there is growing evidence that multiple forms of feedback exist. Studies of visual imagery and working memory indicate the flexible nature of top-down feedback from frontal-parietal areas to early visual areas for maintaining and manipulating visual information about stimuli that are no longer in view. Theories of predictive coding propose that higher visual areas encode feedforward signals according to learned higher order patterns, and that any unexplained components are fed back as residual error signals to lower visual areas for further processing. These feedback error signals may serve to define an image region as more salient, figural, or stronger in apparent contrast. Here, I will discuss both theory and supporting evidence of multiple forms of top-down feedback, and consider how deep learning networks can be used to evaluate the utility of predictive coding models for understanding vision. I will go on to discuss what important questions remain to be addressed regarding the nature of feedback in the visual system.