Hidden layer learns a generative model of a long sequence of spike trains.

<p>(<i>A</i>) The three neurons in layer 1 (Layer 1) are clamped to input spike trains (in this example, a random pattern with spike frequency of 0.2). During the open state, these neurons project (through synaptic weights sampled from a random uniform distribution between -1 and 1) to a much larger hidden layer (Layer 2) which in this example contains thirty neurons. Synaptic state matching in the intra-layer 2 synaptic connections (L2-L2) gives rise to an accurate predictive model within this hidden layer. Synaptic state matching of synapse projections from layer 2 to layer 1 (L2-L1) generates an accurate predictive model of the missing input at layer 1. Layer 1 neurons do not have any internal recurrent connections. (<i>B</i>) top: missing input into layer 1 during the closed state; middle: predicted pattern of missing input created by the projection of hidden layer axons to layer 1 during the closed state. The pattern is a perfect match to the missing input; bottom: the coincident activity of the 30 neurons in the hidden layer during the closed state. SSM parameter choices for both L2-L2 and L2-L1 connections were as follows: potentiation strength (α): 5×10⁻⁴, spike-rate memory (m_s): 200, potentiation memory (m_p): 50, state switching period: (Gaussian, τ_μ = 30, τ_σ = 10), neuron firing threshold (V_t): 0.5, sigmoid sharpness (S): 10, latency range (L): 20.