Fig. 3: Algorithm implementation on the MCA.
From: Domain wall motion-driven magnetic convolutional accelerator
a Conceptual illustration of the STFT. The input signal is segmented and Fourier-transformed. b Schematic of 4-point STFT using a 7-Hall-pad MCA device. Input values \({a}_{n}\) map to \({L}_{D}\), and kernel coefficients \({b}_{k-n}\) to \({W}_{P}\). Sequential domain shifting and AHE readout yield Fourier components \({X}_{n}\). c Magnetic-domain snapshot as domains shift across the 7-Hall-pad device. Scale bar: 14 µm. d Experimental STFT demonstration. (i) Initial segment of \(f\left(t\right)\). (ii) Final segment of \(g\left(t\right)\). (iii) Combined input \(f\left(t\right)+\,g\left(t\right)\). (iv) MCA-generated STFT output showing the temporal evolution of both components. e CNN flow diagram for MNIST handwritten digit recognition. f Implementation of a 3 × 3 convolutional kernel using three MCA devices, each encoding a 1 × 3 kernel via \({W}_{P}\) configurations (gray dots). Pixel intensities map to \({L}_{D}\). Scale bar: 20 µm. g Confusion matrix for the MNIST task, showing 98% average accuracy. h Left: 256 × 256 greyscale ‘Cameraman’ image. Right: Intensity profile along a line cut; x-axis denotes pixel position. i Experimental setup for edge detection using the kernel [1, 0, –1] encoded via \({W}_{P}\), with pixel intensities mapped to \({L}_{D}\). Scale bar: 20 µm. j Edge-detected image produced using the MCA.
