Abstract
Physiological studies have reported that V4 neurons represent curvature and its direction (Carlson et al., 2011). We investigated what controls the construction of the curvature selectivity, and what is necessary for the construction. We consider that sparseness is the key for understanding this issue. To investigate the sparseness in the construction, we applied sparse coding to the activities of model V2 neurons in response to natural images so as to obtain basis functions corresponding to the RFs of V4 neurons. With the sparseness ranged between 0.7-0.8, the curvature selectivity of each basis and their population activity emerged as similar to the physiology. This result indicates that sparseness is sufficient to control the construction of the curvature selectivity. In the model above, the RFs of model V2 neurons consisted of two Gabor filters. Depending on the combination of their phase, some models may represent a surface (e.g., vertically aligned cells with the same phase), but others may not. To investigate whether the surface representation is necessary for the construction of the curvature selectivity, we classified the models into the two categories, and analyzed the dependence of selectivity on the categories. Model neurons that included surface representation yielded curvature selectivity, and the others did not. This result indicates that surface representation is necessary for the construction of the curvature selectivity. Because appropriate sparseness was required for the construction of the selectivity, it is expected that the model V4 neurons with the selectivity should show the appropriate sparseness (0.7-0.8). To confirm this expectation, we compared the lifetime sparseness of the model cells and that of the basis functions. The distributions of lifetime sparseness were, in fact, very similar. These results indicate the crucial role of sparseness and surface representation in the representation of curvature and primitive shape in V4.
Meeting abstract presented at VSS 2014