Fig. 4: Signal processing based on in-memory processing.

a, Description of convolution-based signal processing for different filters (low-/high-pass filters and identity). y, processed signal; x, input signal; h, filter kernel. The kernel is split between its positive and negative components; these values are proportionally transferred to the memory weights. The input signal is simultaneously applied to all the memories and the difference between the output of two columns is the result of the processed signal for a given kernel. b, Comparison of the theoretical kernel weight mapping and the experimental weight transfer into the conductance of the memories. c, Comparison of the fast Fourier transform (FFT) of the simulated and experimental output signals after each kernel.