Abstract
We examined the essence of the receptive-filed properties of V4 neurons from the view point of information, specifically whether the coding of V4 neurons is in a form of sparse representation. The component analyses with the constraint of sparseness were applied to the image patches that simulated the responses of either V1 or V2 neurons for natural images. With the image patches that corresponded to the outputs of V2 neurons, we obtained a set of basis functions that were selective to contours with particular curvatures, which is similar to V4 receptive-fields. The result suggests the significance of sparse coding in the representation of V4 neurons, and the importance of afferent inputs from V2 in the formation of the V4 receptive-filed property.
