Abstract
Earlier neurophysiological studies by Pasupathy and Connor (1999, 2001, 2002) have showed that some V4 neurons are tuned to cue-invariant corner stimuli, suggesting that these neurons might be coding abstract concept of corners. Here, we studied whether and how neurons in early visual cortical areas are tuned to a more general set of edge concepts derived from natural images using statistical clustering techniques. By edge concepts, we mean cluster centers or prototypes of aligned edge contour segments extracted from the human-annotated edge maps of Berkeley segmentation. These concepts, totaling about 200, consist of lines, curves, junctions as well as more complex patterns. Each of these edge contour prototypes is associated with a particular mean image appearance patch where it comes from, and a set of principal components of all the appearance patches associated with the edge segments in that cluster. We presented these edge concepts, their associated appearance patches, and their first principal component patches to the receptive fields of V1, V2 and V3a neurons recorded from awake behaving monkeys using a semi-chronically implanted multi-electrode array. We found that the responses of a majority of the neurons in V1 and V2 to the contour patterns are correlated with their responses to the corresponding appearance patterns or principal component patterns (R =~0.5), suggesting that the neurons are tuned to these edge concepts. Individual neuron's coding of the edge concepts however is sparse in the sense that each will strongly prefer a few edge concepts, but is also distributed in the sense that it also responds moderately to many other concepts. The few V3a neurons recorded however appeared to respond strongly and transiently to the appearance patches but not the contour patches, suggesting that these neurons, with much larger receptive fields, might care more about surface concepts than edge concepts.
Meeting abstract presented at VSS 2014