Fig. 4: MNIST pattern recognition simulation.

a Constituents of a single-layer network for the typical MNIST pattern recognition process in which input neurons (yellow) and output neurons (green) are fully connected by synaptic weight (blue). b Diagram of a crossbar array mapped into ANNs. The input voltages (x_i, grey) are applied to the input neurons of each column (yellow) and are scaled by their own synaptic weights (w_i,j = G_i,j⁺−G_i,j⁻). The output signals are integrated into the form of ∑w_i,jx_i at the output neurons of each row (green). c, d Diagrams of reconstructed ANN architectures for two cases: (i) the case in which the undesired neural pathways (red lines) are not suppressed, and (ii) the fully suppressed case. e, g Reshaped 28 × 28 contour images of the final conductances (G_i,j⁺ and G_i,j⁻) and synaptic weights (w_i,j) corresponding to “3” after 15 training epochs corresponding to S.C = 1 and S.C = 0. f, h Confusion matrices for a classification test of the 10,000 MNIST handwritten digit images after 15 training epochs corresponding to S.C = 1.0 and S.C = 0