Cue-Invariant Geometric Structure of the Population Codes in Macaque V1 and V2

Abstract

AO_SCPLOWBSTRACTC_SCPLOWOur ability to recognize objects and scenes, whether they appear in photographs, cartoons, or simple line drawings, is striking. Studies have shown that infants, members of isolated Stone Age tribes, and non-human primates can readily identify objects from line drawings. These findings suggest that the brain may inherently generate neural representations that align across different rendering cues, enabling abstraction. To test this hypothesis, we investigated the representational invariance of complex patterns of surface boundaries found in natural scenes. We tested whether individual neurons in V1 and V2 of the macaque monkey responded similarly to the presentation of these patterns across different renderings (i.e., as contours, luminance-defined patches, and segments of natural images). We found that individual neurons exhibit some degree of tuning invariance, stronger in V1 than in V2. At the population level, as a means to assess cue-invariant abstract representation, we measured decoding accuracy across cues ( cue-transfer decoding). We found that this decoding is greatly enhanced when a geometric transformation (Procrustes Transformation) is first performed to align the population activities across cues. It is also effective when applied to different populations of neurons within or across visual areas. These results were compared with populations of artificial neurons from models of the ventral visual streams, further indicating that cue-invariance stabilizes with population size. In summary, we found that while individual neurons exhibit some cue-invariance properties, the stability of the population geometry emerges as a more robust candidate for supporting a cue-invariant representation of visual information in the early visual areas. SIGNIFICANT STATEMENTHow can we easily recognize objects and scenes in a wide range of renderings, such as photographs, cartoons, or line drawings? One possibility is that our visual system processes information using an invariant representation. To investigate this hypothesis, we designed a stimulus set made of boundary patterns extracted from natural scenes, and displayed using three distinct renderings. We found that, while the tuning preference of individual V1 and V2 neurons displayed some correlation across renderings, a more robust invariant representation could be achieved when analyzing neural population geometry. Overall, we found that a cue-invariant representation of visual elements in the early visual areas may rest primarily on the geometry of the population responses, rather than individual neurons tuning characteristics.