Spike-timing-dependent plasticity can be rephrased as noisy gradient, which allows to obtain convergence guarantees for the stochastic learning dynamics.