Abstract
Although the representation of space is as fundamental to visual processing as the representation of shape, it has received relatively little attention. Here we develop a neural model of two-dimensional space and examine how the representation is affected by the characteristics of the encoding neural population (RF size, distribution of RF centers, degree of overlap, etc.). Spatial responses of the model neurons in the population were defined by overlapping Gaussian receptive fields. Activating the population with a stimulus at a particular location produced a vector of neural responses characteristic for that location. Moving the stimulus to n locations along the frontoparallel plane produced n response vectors. To recover the geometry of the visual space encoded by the neural population, the set of response vectors was analyzed by multidimensional scaling, followed by a Procrustes transform. The veridicality of the recovered neural spatial representation was quantified by calculating the stress, or normalized square error, between physical space and this recovered neural representation. The modeling found that large receptive fields provide more accurate spatial representations, thus undermining the longstanding idea that large receptive fields in higher levels of the ventral visual pathway are needed to establish position invariant responses. Smaller receptive field diameters degrade and distort the spatial representation. In fact, populations with the smallest receptive field sizes, which are present in early visual areas and, at a single cell level, contain the most precise spatial information, are unable to reconstruct even a topologically consistent rendition of space. Development of this neural model provides a general theoretical framework not only for understanding neurophysiological spatial data, but also for testing how various neuronal parameters affect spatial representation.