Neural Computations in Visual Cortex

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): The goal of this research is to understand the nature of the computations performed by primary visual cortex (V1), and how these calculations are carried out. Even the most basic step in interpreting the visual world -- extracting local features such as lines and edges -- is a difficult computational problem: it must be carried out in the context of cluttered, complex, natural visual scenes; it must be carried out rapidly; and it must be carried out by neural hardware. The generally accepted view is that V1 acts primarily as a feedforward bank of filters, in which feedback and gain controls play a modulatory role. However, models constructed from simple analytically-convenient stimuli provide an incomplete account of responses to natural scenes. Since natural scenes have characteristics that traditional analytic stimuli lack, this observation implies that V1 neurons are sensitive to these distinguishing characteristics, namely, high-order statistics (HOS's). Based on several lines of evidence (including work from the previous funding period and studies in other laboratories), we hypothesize that this sensitivity to HOS's indicates that V1's basic design is that of a strongly recurrent network. In particular, we hypothesize that the characteristics that distinguish a strongly recurrent architecture from a feedforward or modulatory feedback architecture account for V1's ability to extract HOS's. To test these hypotheses, we focus on analyzing V1's responses to stimuli containing HOS's -- because they distinguish among these two contrasting pictures of V1, and because HOS's are precisely the statistical feature that distinguishes natural scenes from traditional analytic stimuli. In Aim 1, we determine the extent of sensitivity of V1 neurons to HOS's, explicitly studying both artificially- constructed stimuli and stimuli derived from natural scenes. In Aim 2, we determine whether dynamic formation of neural assemblies underlies the extraction of HOS's, by analyzing the statistics of multineuronal firing patterns. If successful, this work will provide fundamental insights into the design principles of V1, including how it exploits general features of cortical architecture to carry out the calculations necessary for vision, how sparse representations arise, and the functional significance of cortical neural "noise."