Figure 3.

Neural Sampling in an RBM consisting of 10 stochastic I&F neurons, with five neurons in each layer. Each neuron is associated to a binary random variable which take values 1 during a refractory period τ_r after the neuron has spiked (gray shadings). The variables are sampled at 1 kHz to produce binary vectors that correspond to samples of the joint distribution p(z). In this figure, only the membrane potential and the samples produced by the first five neurons are shown. The vectors inside the brackets are example samples of the marginalized distribution p(z₁, z₂, z₃, z₄, z₅) produced at the time indicated by the vertical lines. In the RBM, there are no recurrent connections within a layer.