Fig. 6: In-memory computing of a pattern classification task within a passive crossbar array.

a, b The optical image (a) and schematic diagram (b) of a 40 × 40 passive crossbar array based on FFD devices. c The distribution of negative coercive voltage and positive coercive voltage obtained from transient I–V curves in Fig. S16. The red lines indicate the fit by gaussian distribution. d Statistics of resistive switching ratio at a read voltage of 3 V in the 40 × 40 passive crossbar array. e A schematic diagram of an artificial neural network (ANN) for clarifying three 4 × 4-pixel images. f The schematic diagram indicating how the hardware to implement the ANN in (e). g The schematic diagram indicating how the image are encoded to input into the hardware ANN. Each weight is encoded by conductance difference between neighboring pair units. h–j The current difference (I_i = I_i⁺-I_i^-) collected from columns when images of “L” (h) “u” (i) and “n” (j) are inputted into the 16 × 6 hardware ANN. The patterns recognition is implemented with the value of I_i standing for the target i image being the biggest. k When zero, one, and two pixels are randomly flipped, and recognition accuracy is 100%, 50%, and 37.5%, respectively.