Figure 2.

Schematic of the simulated neural network: Positive and negative voltage pulses are applied to the input of the network, representing the intensity of the individual values of the 28 × 28 pixel MNIST image. Stochastic coding of the input data has been implemented by Poisson-spiking input neurons (blue circles) with firing rates proportional to the intensity of the corresponding pixel of the input pattern. Red circles are leaky integrate-and-fire output neurons (LIF), which are laterally coupled in an inhibitory winner-takes-it-all network (WTA). The individual memristive devices are arranged in a crossbar structure. A local STDP-based learning rule has been implemented using the defined I–V nonlinearity of the memristive devices: Only the overlap (association) of pre- pulses and post-pulses leads to an increase (potentiation) or decrease (depression) of the device conductance.