A theory of orientation selectivity emerging from randomly sampling the visual field

Abstract

AbstractSelectivity for the orientation (OS) of moving gratings is also observed in the primary visual cortex (V1) of species which lack an orientation map, like mice and rats. However, the mechanism underlying the emergence of orientation selectivity in these cases is still mysterious. In the visual pathway, signals from the retina are conveyed to V1 via relay cells in the dorsal lateral geniculate nucleus (dLGN). Here, we explore a computational model of orientation processing in dLGN and V1, where orientation selectivity emerges from random thalamocortical projections and investigate the underlying neuronal mechanism. First we demonstrate that the temporal mean (F0) of the thalamic input to cortical neurons does not depend on orientation, while the amplitude of the temporal modulations (F1) induced by the stimulus grating is indeed tuned to stimulus orientation. Second we show that the orientation tuning in the amplitudes of thalamic input can be transformed into orientation selectivity of V1 neurons by a random network, exploiting the nonlinear input-output transfer function of individual neurons. The untuned excitatory component is largely cancelled by the inhibitory component, such that the operating point is stabilized in the nonlinear regime of the transfer function. In our model, the orientation preference of V1 neurons is sensitive to the spatial frequency of the visual stimulus, very similar to what is found in neuronal recordings. Finally, we describe a non-linear firing rate model that accounts, among other things, for the F1-to-F0 transfer of oscillatory input to neurons.Author summaryOrientation selective neurons are widely observed in mammalian primary visual cortex. Previous theoretical work has argued that contrast-invariant orientation selectivity arises in inhibition-dominated random recurrent networks by attenuating the uninformative component and amplifying a weak orientation bias in the input. Based on a computational model of the thalamocortical pathway, we show that discrete isotropic sampling of the visual field can generate the necessary orientation bias by uneven distribution (random symmetry breaking). This bias is sufficient to induce strong orientation selectivity in a cortical network by exploiting the non-linear transfer of individual neurons. The model exhibits not only biologically plausible behavior of neuronal network, but also suggests a mechanism for the emergence of orientation selectivity in V1. A new firing rate model allows us to determine the output firing rate for oscillatory input without solving the time-dependent Fokker-Planck equation.