CONICET | Buscador de Institutos y Recursos Humanos

The neural sampling hypothesis suggests that the brain?s sensory processing can be understood as implementing a sampling process that computes the posterior belief over latent variables given a sensory observation [1]. Prior studies have examined this hypothesis in the context of the primary visual cortex (V1), and almost all of them have assumed V1 neurons to represent the continuous intensity of Gabor-shaped features either by firing rates [2-4], or membrane potentials [5-7]. All of these models either implemented known sampling algorithms or simply ignored dynamics (though see [8]). Most recently, Echeveste et al. [9] showed that a rate-based model optimized for fast sampling exhibited rich dynamics including marked stimulus-dependent transient responses and gamma oscillations.It remains unclear whether these dynamic features observed in models of continuous variables would still emerge in the case of binary latents. However, learning in Gaussian binary models of natural images also gives rise to Gabor-shaped receptive fields [10]. Gibbs-sampling based inference in binary models has been shown to be feasible using spiking networks [11,12] and using leaky integrate-and-fire neurons [13], which yielded near contrast-invariant tuning curves and realistic spiking statistics, interpretable as probabilistic population codes (PPCs, [14]).Here, we optimized the recurrent connectivity of small networks consisting of binary excitatory and inhibitory neurons (Fig. 1B) that obey Dale?s Law to maximize the speed of sampling-based inference in a generative model of retinal images (Fig. 1A). We found sampling was fastest for non-symmetric connectivity structures that implied oscillatory dynamics [15] with performance between Gibbs sampling and independent sampling (Fig. 1C). Importantly, unlike Gibbs sampling and independent sampling, neural responses in our network presented strong oscillations (Fig. 1D), as observed in the cortex.