Approximation quality of networks with different refractory mechanisms.

Mean and standard deviation of the Kullback-Leibler divergence between reference Boltzmann distributions and neural sampling approximations for three different neuron models (corresponding to columns) and three different values for the reference distribution hyperparameter (corresponding to rows). The parameter controls the standard deviation of the weights of the reference distributions . In case of very strong synaptic interactions (leading to sharply peaked distributions, ) the approximation quality of the spiking network degrades, if the neurons feature a relative refractory mechanism. The data was computed from 100 randomly generated Boltzmann distributions and their neural approximations for each value of .