V2 builds a generalizable texture representation

Decades of work has attempted to explain how the representation of visual information is transformed across stages of the cortical processing stream. In some cases, these transformations have been elucidated by insights into how the receptive field properties of individual neurons differ across areas. But the view provided by single neuron selectivity has struggled to explain the transformations occurring from early to mid-level cortex. An alternative, complementary approach is to study the representational geometry of neuronal populations. Though derived from individual neuronal selectivity, population geometry can reveal encoding strategies difficult to infer from single neurons. Recent work in artificial networks has shown the trade-offs of different visual representational geometries. For instance, discriminability is favored by high-dimensional representations; generalizability is favored by low-dimensional, systematic embedding of image statistics. To understand the representational strategy in the visual cortex, we compared the population geometry for synthetic, naturalistic textures in primary visual cortex (V1) and area V2 of macaque monkeys. We simultaneously measured neuronal population activity in both areas, and determined their sparsity, dimensionality, discriminability, latent embedding, and generalizability performance, for an ensemble of textures. V2 populations showed evidence of an emerging latent representation of image statistics, affording better generalizability in predicting responses to new textures. This latent representation also allowed for improved discriminability between textures. Analysis of pre-trained deep net (AlexNet and VGG-16) responses to the same stimulus ensemble did not replicate the representational geometry observed in cortex. We conclude that there is a shift in the representational geometry between V1 and V2 to a latent representation of image statistics. Our results show that assessing representational geometry can provide important insights into the transformations that occur across successive stages of visual processing.