Fig. 5: Robustness of the PD-RRAM array for image recognition.

a The measured dark current of the I_out under the short-circuit condition. b Current–voltage characteristic curve of the redesigned RRAM under DC sweeping voltage between −5 V and 7 V with a SET compliance current (I_CC) of 500 μA. c Schematic illustration of the PD-RRAM array with a differential pair of RRAMs to realize signed weights. The MAC operations of positive weights (R⁺) and negative weights (R⁻) are performed at t₀ and \({t_{0}}+\varDelta{t}\,({\varDelta{t}}\,<\,1\,{\upmu{\rm{s}}})\), respectively, and are then subtracted from each other by peripheral circuits. d Accuracy of the classifier training for different weight standard deviations. The classifier has 784 pixels and is used to distinguish ‘5’, trained on the MNIST database of handwritten digits. e Accuracy of the classifier training for different failed pixel percentages. Here the classifier sets the weights of random pixels to 0, trained as d