784 Inputs 28 × 28 × 1 64 Nodes 64 Nodes 10 Nodes 10 Nodes 10 Output Accuracy of MNIST 1 9 MACU Input Layer Fully Connected Layer 1 Relu Layer (a) Clean Speech 1 0.5 -0.5 -1 01 23 Noisy Speech 1 0.5 -0.5 -1 01 23 Denoised Speech 1 0.5 -0.5 -1 01 23 Time (s) (c) Original Gray Image Standard Edge Detection Edge Detection by MACU 2.5 2 1.5 1 0.5 0123 4 Frequency (KHz) 2.5 2 1.5 1 0.5 0123 4 Denoised Speech -50 -100 -150 2.5 2 1.5 1 0.5 01 2 Noisy Speech -50 -100 -150 34 Clean Speech -50 -100 -150 Fully Connected Layer 2 Softmax Layer Output Layer 0.8 0.6 0.4 0.2 6 7 8 ADC Resolution 9 2 (b) 3 4 Standard Accuracy MACUs Accuracy 5 6 Weight Resolution (d) FIGURE 19: The MACUs implementation. (a) A five-layer neural network for the Modified National Institute of Standards and Technology (MNIST) data set classification, (b) testing accuracy of the data set, (c) speech denoising with MACUs, and (d) edge detection with MACUs [36]. ReLu: rectified linear unit; ADC: analog-to-digital converter. IEEE SOLID-STATE CIRCUITS MAGAZINE FALL 2021 41 Time (s) Time (s) Time (s) Accuracy Power/Frequency (dB/Hz) Power/Frequency (dB/Hz) Power/Frequency (dB/Hz)