Fig 1. Basic properties of the network match experimental findings.

(a) The Self-Organizing Recurrent Neural Network (SORN) consists of recurrently connected excitatory (blue) and inhibitory (red) deterministic McCulloch & Pitts threshold neurons. Each input letter (black boxes) stimulates an excitatory subpopulation. The excitatory recurrent connections are shaped by spike-timing dependent plasticity and synaptic normalization. The excitatory thresholds are regulated by intrinsic plasticity (see Methods for details). (b) Raster plot of spontaneous activity (no external input) after stimulating the network with ten randomly alternating letters during plasticity. (c) The inter-spike-interval (ISI) distribution of a randomly selected neuron during spontaneous activity is well-fitted by an exponential apart from very small ISIs. (c, inset) The distribution of coefficients of variation (CVs) of the ISIs clusters around one, as expected for exponential ISI distributions, compatible with the experimentally observed Poisson-like spiking [10, 40]. (d) The fraction of excitatory-to-excitatory connections converges to a stable fraction. (e) Individual weights fluctuate despite the global convergence as observed experimentally [41]. (f) After self-organization, i.e. at the end of (d), the binned distribution of excitatory-to-excitatory synaptic weights (dots) is well fit by a lognormal distribution (solid line, cp., e.g., [42]).