Supplemental Material to Network Plasticity as Bayesian Inference

where μ1 = 0.3, μ2 = 0.9, σ1 = 0.1, σ2 = 0.2 and c = 0.3. In Fig. 1D we used a prior pS(θ) = pS(θ1)pS(θ2), with pS(θi) given by a normal distribution (μ = 0.3, σ = 0.35). A learning rate of η = 0.005 was used to sampled trajectories which had a length of 50 and 300 time steps in Fig. 1C and F, respectively. In Fig. 1F the time-discrete version of the synaptic sampling algorithm (7) was used, with N = T = 1. In Fig. 1C the same dynamics were used, but the diffusion term and the contribution of the prior ∂ ∂θi log pS(θ) were set to zero.