How can cells in the anterior medial face patch be viewpoint invariant?

Abstract

IntroductionIn a recent paper, Freiwald and Tsao [1] found evidence that the responses of cells in the macaque anterior medial (AM) face patch are invariant to significant changes in viewpoint. The monkey subjects had no prior experience with the individuals depicted in the stimuli and were never given an opportunity to view the same individual from different viewpoints sequentially. These results cannot be explained by a mechanism based only on the temporal association of experienced views.Affine transformations such as translation and scaling can be computed independently of the particular object and are thus generic. On the other hand, non-generic transformations including changes in viewpoint and illumination depend on the 3D structure or material properties of the objects [2].Within a restricted class of objects having a similar 3D structure -like faces -it may be possible to have model cells which are identity-specific and invariant to non-generic transformations using class-specific features [3]. Freiwald and Tsao's finding of a population of cells that appear to be invariant to changes in viewpoint could constitute evidence that the brain is employing face-specific features for this purpose. Simulation ProcedureWe simulate a same-different task by presenting images that contain either the target object or a distractor object on a neutral background. The object may appear at a range of viewpoints. We trained a simple correlation-based classifier on a single view of the target object presented frontally (rotation of 0). The task was for the classifier to rank transformed versions of the target object as being more similar to the trained view than to distractors and rotated distractors. We tested the model with increasing ranges over which we allowed targets and distractors to rotate. The range of viewpoints tested in each condition is plotted as the abscissa in the following figures. The classifier only used the output of the model (C2, S3 or C3) and no other information.

Representation Learning in Sensory Cortex: A Theory

“…In fact, the same localization condition of Equation 4 suggests Gabor-like templates for generic images in the first layers of a hierarchical architecture and specific tuned templates for the last stages of the hierarchy, since class specific modules are needed with each containing highly specific templates, and thus highly tuned cells. This is consistent with the architecture of the ventral stream and the the existence of class-specific modules in primate cortex such as a face module and a body module ( [30], [128], [76], [63], [52], [58], [28]). We saw in section IIF that areas in the hierarchy up to V4 and/or PIT provide signatures for larger parts or full objects.…”