Task and model description. A, Left, Stimuli-evoked activity patterns v (ticks) with probabilities given by a target distribution P*(v) (curves). Middle, In the absence of stimuli the network spontaneously generates activity patterns x = (v, h) distributed according to a model distribution P_w0 (x) with synaptic strength parameters w₀. Right, With learning, the distribution of spontaneous activity patterns P_w(v) approaches the target distribution. Hidden neurons (in gray) help to support the desired activity patterns in the visible neurons. B, Neural dynamics define the model distribution P_w(x) trough a spike response kernel ε, an adaptation kernel κ, and the spike probability ρ. Minimizing the divergence (P*(v)‖P_w(v)) from the target to the model distribution leads to a plasticity rule for the synaptic strengths w that is matched to the recall dynamics.