Abstract
Marr (1981) proposed that generalized cones (GCs)-the shapes or volumes produced by sweeping a cross section along an axis-provide a general formalism for describing shapes corresponding to simple parts of an object. Differences between GCs are described in terms of simple properties of the generating function, such as the curvature of the axis or changes in the size of the cross section (which would produce changes in the parallelism or curvature of the sides). To what extent can the tuning of inferior temporal (IT) neurons to variations in shape be described as reflecting sensitivity to these shape dimensions? We recorded the responses of 40 IT neurons from 2 monkeys while they viewed simple line drawings corresponding to 2D projections of a GC (which could be produced by a 1D cross section, viz., a line). The drawings could be arranged into 2 low-dimensional shape spaces. The first set consisted of variations of a box that could become progressively more tapered or curved. The curvature dimensions were the curvature of the main axis and positive or negative curvature of the sides. The second shape space had a triangle at its center and included only the 3 curvature dimensions. Some of these dimensions may be expected (based on pixel- and wavelet-based differences) to be either entangled (e.g. curvature of the main axis and curvature of the sides) or to extend each other (e.g. negative and positive curvature of the sides). Nevertheless, multidimensional scaling analysis of the neural activity resulted in approximately orthogonal dimensions in both spaces. Also, different neurons were sensitive to different dimensions. These results may provide the neural basis for the psychophysical ability to selectively respond to one dimension while ignoring variations in another dimension. The average neural response to our shapes increased with the distance from the center, suggesting a preference for shapes with extreme rather than intermediate values on these dimensions.