Extended Data Fig. 4: Ten-digits MNIST classification using kernel-based convolution neural network fit on our device.
From: Synthetic-domain computing and neural networks using lithium niobate integrated nonlinear phononics
a, Measured spectrum of the input and output of the first layer. The input includes the image pixels (28×28) and 128 non-zero parameters (8 kernels, each has 16 non-zero parameters) with df = 50 Hz and f0 = 1,022.92 MHz. The output includes (2×784-1 = 1,567) elements of quadratic operation of input image and 8 convoluted features (each has 847 parameters). The gray curve at the second-order frequency is self-convolutions of kernels and will be multiplied by zeros in the fully connected layer. The first 8,480 elements are fed into a digital computer to multiply with a 10 × 8,480 matrix. b, Confusion matrices of the calculated and experimentally measured inference results of the first 1,000 validation images, showing similar performance of inference with a calculated (partially experimental) accuracy of 95.1% (95.0%). c, NMSE of measured results of the first layer compared to calculated results.
