Abstract
The origin of orientation selectivity in primary visual cortex is still a hotly debated issue. Two main classes of models — the sharpening and no sharpening models — have been proposed. In the no sharpening model, orientation selectivity is the result of the convergence of LGN afferences onto cortical cells. By contrast, in the sharpening model, the pooled input from the LGN is broadly tuned but subsequently sharpened through cortical lateral interactions. Because these models can generate similar tuning curves, they are commonly thought to produce similar codes for orientation. We show here that this is typically not the case. Even when the tuning curves and spike train variability of individual neurons are the same in the two models, the pairwise correlations are markedly different, and the no sharpening model conveys far more information about orientation than the sharpening model. Moreover, the majority of the information in the sharpening model is conveyed by correlations, making the sharpening model particularly inefficient for learning and computation. These results demonstrate that sharpening through lateral connections is not always as beneficial as generally believed.